1. Abstract

2. The Cognitive Environment

   • Ukraine

   • Iran, Israel, and Cyber-Enabled Influence

3. Attention, Epistemic Trust, and Resilience

   • Epistemic Trust Degradation

4. From Behavior to State Estimation

   • What Machines Can Actually Infer

5. Altered States and Sleep

   • Gateway

   • Tetris, Targeted Memory Reactivation, and Dream Incubation

6. The Closed Loop: An Old Principle at a New Scale

   • Closed-Loop Error Amplification

7. Delegating Control Can Increase or Reduce Autonomy

   • Autonomy-Enhancing Delegation

   • Autonomy-Reducing Delegation

   • Who Controls the Objective?

8. Consent Cannot Be a One-Time Click

9. Cognitive Autonomy Is Unevenly Distributed

10. Economics May Stop a Technology More Effectively Than Physics

11. Global Asymmetry

12. Responsibility in Distributed Decision Systems

13. Extreme Conditions

14. State of the Art: What We Still Cannot Do

15. The Operator May Be Distributed

16. Meta-Evolution Interpretation

• Tool Augmentation vs. Adaptive Control Migration

17. Future Horizon: From Brain-as-a-Service to Cognition-as-a-Service

• Near-Term: AI That Adapts to the User’s State

• Cognition-as-a-Service: Cognitive Resources on Demand

• Hybrid Wake-Sleep Cognition

• AI-Orchestrated Cognitive Clusters

• The Economic Test

• Who Owns the Purpose of Cognitive Work?

18. The Right to One’s Own Operator?

19. Conclusions

20. References

Abstract

Digital technologies are becoming increasingly involved in the process that leads from information to decision. Algorithms select content, recommendation systems determine visibility, influence operations attempt to shape the interpretation of events, and generative AI sharply reduces the cost of producing text, images, audio, and video. At the same time, another technological layer is developing: brain-computer interfaces, wearable EEG, neuroimaging, physiological sensing, targeted memory reactivation, experiments involving communication during sleep, and adaptive neural stimulation.

These technologies emerge from different fields and operate under very different constraints. Their existence does not demonstrate the emergence of a unified architecture for controlling the human mind. It does, however, justify a more precise question:

Which parts of the process leading from information to decision are becoming technologically observable, modifiable, or delegated to external systems, and when does such delegation extend human agency rather than redistribute control away from the person?

The term human cognitive operator is used here as an analytical shorthand. It refers to the distributed set of processes involved in selecting information, evaluating credibility, integrating experience, updating representations of the environment, and choosing action. It does not imply the existence of a separate neural module or a new neuroscientific construct.

The essay begins with existing evidence concerning cognitive warfare, attention dynamics, neural decoding, sleep research, and closed-loop systems. It then turns to autonomy, economics, consent, responsibility, and asymmetries of power. Meta-Evolution is introduced later as an interpretive framework capable of generating a more specific question about the location of control. The final section is explicitly technological forecasting and extends the discussion toward on-demand clustering of human cognitive resources, during both wakefulness and sleep.

1. The Cognitive Environment

Influencing an opponent’s perception is one of the oldest components of conflict. Propaganda, deception, manipulation of intentions, and psychological warfare existed long before computers.

Four parameters have changed substantially:

scale, speed, personalization, and feedback.

In a digital environment, a message can reach millions of people, audience reactions can be measured almost immediately, multiple variants can be tested, and subsequent communication can be adjusted according to observed responses.

NATO Allied Command Transformation has developed the concept of cognitive warfare around competition for cognitive advantage. NATO’s Science & Technology Organization treats cognition as a cross-cutting dimension of effects rather than as a formally established warfighting domain equivalent to land, sea, air, space, or cyberspace. In 2026, the NATO Defense College devoted its Conference of Commandants to The Mind as a Battlespace and the problem of cognitive resilience.

This represents a functional shift in strategic attention.

The relevant target may lie within the process connecting an event to an action:

event → information → selection → interpretation → decision → behavior

Ukraine

The Russia-Ukraine war demonstrates the scale of the contemporary influence environment. A 2026 analysis by the European External Action Service and Ukraine’s Center for Countering Disinformation identified approximately 244,000 publications generating around 1.39 billion views in campaigns related to Ukraine’s integration with the European Union.

The report described methods including:

• AI-assisted mass production of content,

• cross-platform amplification,

• information laundering,

• reuse of authentic statements outside their original context,

• manufacturing the appearance of broader consensus.

The significance of such activity does not depend on an assumption that audiences automatically believe false information.

Every additional claim may require attention.

Contradictory claims require comparison.

Fabricated evidence requires verification.

Authentic evidence may be challenged as synthetic.

The cognitive and institutional cost of assessment is borne by those attempting to maintain a reliable model of events.

Iran, Israel, and cyber-enabled influence

Microsoft Threat Intelligence has documented Iranian operations combining cyberattacks, narratives about their effects, fabricated online personas, and social amplification.

In such a configuration, a technical event becomes the beginning of a cognitive process:

Different layers of the operation can reinforce one another.

2. Attention, Epistemic Trust, and Resilience

Herbert Simon observed in 1971 that information abundance creates scarcity in the resource information consumes: human attention.

Digital technology has greatly increased the asymmetry between information production and the biological time available to process it.

Lorenz-Spreen and colleagues found evidence of accelerating collective-attention dynamics. Topics rise to prominence more rapidly and disappear from public attention more quickly.

Rathje, Van Bavel, and van der Linden showed a strong relationship between language directed at political out-groups and social-media engagement. McLoughlin and colleagues found that misinformation associated with moral outrage can have a particularly strong diffusion advantage.

This does not imply universal cognitive helplessness.

A large meta-analysis by Pfänder and Altay, covering 194,438 participants in 40 countries, found that people generally distinguish true news from false news with substantial accuracy.

That constraint matters.

Epistemic trust degradation

The problem can emerge before false information is fully accepted.

Three qualitatively different situations can be distinguished descriptively:

“This is true.”
A false claim is accepted.

“I do not know whether this is true.”
Authentic evidence loses part of its credibility.

“I no longer know whether anyone can reliably determine what is true.”
Trust in the verification process itself begins to deteriorate.

The third, limiting condition can be described with the journalistic label truth collapse.

This is not proposed here as an established psychological construct, nor as a claim that contemporary societies have already entered such a state.

A more defensible object of research is epistemic trust degradation.

Generative AI introduces a particular symmetry:

synthetic evidence can appear authentic, while authentic evidence can be dismissed as synthetic.

At the same time, cognition is adaptive.

Research on prebunking and psychological inoculation shows that people can learn to recognize some manipulation techniques. Resilience, however, is unlikely to remain static. Manipulative systems also adapt.

The resulting process may resemble an arms race:

Cognitive resilience therefore deserves to be treated dynamically.

3. From Behavior to State Estimation

The internet can infer a great deal about a person without measuring the brain.

Clicks, language, location, search history, social networks, reaction times, and viewing patterns already create rich behavioral profiles.

Neurotechnology opens an additional channel.

In 2023, Tang, LeBel, Jain, and Huth demonstrated reconstruction of semantic aspects of continuous language from fMRI recordings.

The result was impressive, but required:

• extensive individual calibration,

• participant cooperation,

• a stationary fMRI scanner,

• large quantities of training data.

The researchers also showed that deliberate mental strategies could interfere with decoding.

In the same year, Willett and colleagues achieved high-performance decoding of attempted speech in a person with paralysis using implanted electrodes.

This was a major clinical achievement.

It also illustrates the relationship between signal quality and invasiveness.

A different path is represented by wearable systems. Kaveh and colleagues demonstrated drowsiness classification using in-ear EEG.

What follows from this?

Under particular conditions, machines can already:

• classify some states,

• decode selected motor intentions,

• reconstruct limited semantic information,

• estimate drowsiness,

• detect activity patterns associated with particular tasks.

This remains far removed from unrestricted “mind reading.”

The important shift

Traditional systems primarily observed:

what the person did.

A new generation of sensors can increasingly attempt to estimate:

what state the person is in before acting.

That is a qualitatively different category of information.

4. Altered States and Sleep

The history of attempts to technologically influence cognitive states includes both serious research and highly speculative claims.

Gateway

Analysis and Assessment of Gateway Process, written in 1983 by Lt. Col. Wayne M. McDonnell of U.S. Army Intelligence and Security Command, examined the Gateway program developed by the Monroe Institute.

The document was later preserved in the CIA’s declassified archive, which explains the popular label “CIA Gateway Report.”

Institutional interest does not validate claims concerning remote viewing, out-of-body experiences, or related paranormal phenomena.

Gateway has a more modest and more useful historical significance:

It represents an early attempt to ask whether cognitive states can be systematically shifted through controlled stimulation and training.

Contemporary neuroscience investigates parts of that question using far more rigorous methods.

Tetris, TMR, and dream incubation

Stickgold and colleagues showed that experiences associated with playing Tetris can reappear in hypnagogic imagery during sleep onset.

Targeted Memory Reactivation can, under specific conditions, reactivate previously associated material during sleep. A meta-analysis by Hu and colleagues found an overall effect of TMR on memory consolidation, alongside substantial dependence on protocol and experimental conditions.

Dormio and the Targeted Dream Incubation work of Horowitz and colleagues showed that cues delivered during hypnagogia can increase the likelihood that selected themes appear in subsequent dream reports.

Konkoly and colleagues demonstrated limited two-way communication with some lucid dreamers during REM sleep.

The most defensible common conclusion is:

Sleep does not represent complete isolation from the external environment, but access to the sleeping cognitive system is selective, variable, and strongly dependent on conditions.

TMR does not establish the ability to implant arbitrary beliefs.

Interactive dreaming does not provide an open communication channel to every sleeping person.

Dormio does not demonstrate programmable dreams.

Those boundaries are central to any responsible interpretation.

5. The Closed Loop: An Old Principle at a New Scale

Feedback control is not an invention of the AI era.

Thermostats use feedback.

Neurofeedback has existed for decades.

Adaptive procedures have long been used in psychophysics.

The potentially important change lies in combining:

continuous sensing + multimodality + personalization + AI + automatic adaptation

Adaptive deep brain stimulation offers a real-world example.

In such systems, ongoing neural signals can influence the parameters of subsequent stimulation.

The architecture is straightforward:

measure → estimate state → select intervention → measure again

AI can potentially increase the number of signals integrated and the frequency with which the intervention changes.

It can also create new classes of failure.

Closed-loop error amplification

Consider the sequence:

incorrect measurement
→ incorrect state estimate
→ inappropriate intervention
→ altered human response
→ response interpreted as confirmation of the model.

Feedback can amplify error as well as correction.

This matters in clinical, military, educational, and commercial applications.

6. Delegating Control Can Increase or Reduce Autonomy

This is one of the central distinctions in the entire discussion.

Technological delegation does not automatically represent a loss of autonomy.

A brain-computer interface that restores communication to a person with paralysis can dramatically increase agency.

Adaptive DBS may restore functional capacity.

Decision support may improve performance for a physician in a critical situation.

AI can increase a person’s ability to analyze problems that would otherwise exceed available cognitive resources.

The central question therefore becomes:

Does delegation increase a person’s ability to pursue their own objectives, or does it shift effective control toward a system whose objectives are partially different?

Two broad cases can be distinguished descriptively.

Autonomy-enhancing delegation

Technology increases a person’s capacity to act in accordance with their own goals.

Autonomy-reducing delegation

A system acquires substantial influence over information or decisions while one or more of the following conditions arise:

• objective misalignment,

• limited transparency,

• weak ability to refuse,

• dependence on the system.

This leads to a deeper question:

Who controls the objective?

The same technology can have a radically different ethical structure depending on the relationship in which it is deployed.

7. Consent Cannot Be a One-Time Click

Adaptive systems change.

They learn.

They may expand the range of data they collect.

They may change how they estimate state.

They may alter intervention strategies.

Consent granted at the beginning may therefore fail to describe the system operating a year later.

This makes the idea of dynamic consent particularly relevant to neurotechnology.

Meaningful consent may increasingly require:

• visibility into what is currently being measured,

• notification of new uses of the data,

• control over particular functions,

• the ability to withdraw authorization,

• control over models derived from personal data.

A particularly difficult question follows:

What does deletion mean once a model has already learned a representation of the individual?

This problem extends beyond conventional management of stored personal records.

8. Cognitive Autonomy Is Unevenly Distributed

The risks and benefits of the same technology can differ significantly across populations.

Relevant factors may include:

• age,

• capacity for informed consent,

• dependence on assistive technology,

• cognitive impairment,

• neurological condition,

• availability of alternative forms of communication.

A child using an adaptive educational system occupies a different relationship to the technology than an adult professional.

A person for whom a BCI provides the only effective communication channel has a very different practical ability to “turn the system off” than a user of a commercial application.

Cognitive autonomy cannot therefore be analyzed as if every user were equally independent, equally informed, and equally able to withdraw.

9. Economics May Stop a Technology More Effectively Than Physics

Technical possibility is insufficient.

A system must also be useful and economically rational.

This leads to a simple question:

Why use neurotechnology if behavioral data already works?

A platform possessing:

• interaction history,

• purchasing behavior,

• location data,

• user-generated text,

• social-network information,

• device telemetry,

can already infer a great deal.

Adding EEG becomes attractive only if it produces sufficient additional value.

This creates a strong filter for technological forecasts.

A technology may be:

possible but unnecessary.

It may be:

effective but too expensive.

It may be:

accurate but too dependent on calibration to scale.

It may be:

clinically valuable but commercially irrational.

The same test should be applied to every speculative architecture discussed below.

10. Global Asymmetry

These technologies will not develop evenly across the world.

Countries and corporations differ in access to:

• computing infrastructure,

• medical technology,

• regulatory capacity,

• military resources,

• capital.

The ability to protect neural data and enforce cognitive rights is also unevenly distributed.

This raises several questions.

Who sets the standards?

Where are the data stored?

Who owns models derived from those data?

Which jurisdictions can meaningfully regulate their use?

What protection exists for individuals operating under weaker institutional safeguards?

UNESCO adopted its Recommendation on the Ethics of Neurotechnology in 2025, addressing mental privacy, autonomy, freedom of thought, and the protection of neural data.

A global normative standard does not automatically create global enforcement capacity.

11. Responsibility in Distributed Decision Systems

Closed-loop systems complicate responsibility.

If harm emerges from the interaction of:

a sensor,

an AI model,

an intervention,

a clinical decision,

a user’s reaction,

and subsequent automatic adaptation,

responsibility may no longer have a single obvious location.

This creates a distributed responsibility problem.

Medicine already contains legal frameworks for clinical and product liability.

Military systems may create more difficult cases.

If a commander acts under the strong influence of an AI prediction system, an important boundary appears between:

decision support

and

de facto decision generation.

This essay does not attempt to create a new liability doctrine.

The relevant point is narrower: distributed decision-making can also distribute causal responsibility in ways that existing institutional categories may handle poorly.

12. Extreme Conditions

Laboratory findings should not be transferred automatically to battlefield conditions.

A soldier or commander may operate under:

• sleep deprivation,

• intense time pressure,

• threat to life,

• sensory overload,

• severe stress.

Attention, risk evaluation, and information filtering may operate differently in such conditions than in a controlled online experiment.

This is one of the major limitations of much current discussion surrounding cognitive warfare.

The scale of an information campaign does not itself reveal its effect on a particular operational decision.

13. State of the Art: What We Still Cannot Do

At this point the argument needs a deliberate brake.

Neural decoding

There is no universal device capable of reading a person’s private stream of thought.

fMRI

It provides rich information but requires large, specialized, stationary equipment.

Invasive BCI

It can provide high-quality signals but requires surgery.

Wearable EEG

It is more mobile but provides noisier and more limited signals.

Generalization

A system calibrated on one individual may perform poorly on another.

TMR

Its effects depend on material, protocol, sleep stage, and participant.

Dream interaction

It occurs under particular conditions and does not constitute a universal communication channel.

Closed-loop neurotechnology

It works in selected clinical applications, while biomarker stability, artifacts, and long-term performance remain important challenges.

AI

A model can misclassify state or optimize the wrong proxy.

Integration

The existence of individual components does not demonstrate that their combination will be reliable, scalable, safe, or economically rational.

That final point is crucial.

This essay describes potential points of technological intersection, not a demonstrated convergence toward a unified cognitive-control system.

14. The Operator May Be Distributed

The human cognitive operator should not be imagined as something enclosed entirely inside the brain.

The extended-mind framework associated with Clark and Chalmers, together with the broader tradition of distributed cognition, emphasizes that human cognitive processes can involve tools, language, external memory, other people, and information environments.

Many contemporary problems are already solved by systems resembling:

human + smartphone + search engine + AI + external memory + social network

Technology does not need to enter the brain directly to influence decision formation.

It can organize information order.

It can structure external memory.

It can prioritize options.

It can reduce the visibility of alternatives.

This leads to a broader question of alignment of control.

Meta-Evolution Interpretation

Everything discussed so far can be studied without Meta-Evolution.

Cognitive warfare, neural decoding, brain-computer interfaces, sleep research, adaptive stimulation, and AI each possess their own scientific literatures and validation criteria.

Meta-Evolution does not provide empirical confirmation of those phenomena.

It can, however, introduce a useful organizational distinction.

In Meta-Evolution v1.97, Adaptive Control Migration concerns the redistribution of the decision/control dimension across organizational levels.

Applied cautiously to human-AI systems, this generates the following question:

Is a technological system merely increasing the capabilities of the same decision-making locus, or are some decision/control functions being redistributed across the human, the technology, and the larger hybrid organization?

This distinction goes beyond the simple fact that a tool is being used.

Case A

AI generates several options.

The person has independent sources, evaluates the alternatives, and retains meaningful freedom to choose.

This may represent strong augmentation without sufficient grounds to infer migration of control.

Case B

AI:

• selects most incoming information,

• maintains contextual memory,

• estimates the user’s state,

• generates recommendations,

• dynamically adapts subsequent interventions,

• while the user has little practical access to independent verification.

A different organizational question now appears:

Has the control profile of the system changed?

This essay does not answer that question.

A formal application of Adaptive Control Migration would require the operationalization and validation procedures specified by Meta-Evolution.

Here ME functions as a generator of a research question:

How can tool augmentation be distinguished from a genuine redistribution of adaptive control?

Hoffman, J. Meta-Evolution: An Independent Research Program on Adaptive Information and Higher-Order Evolution. Version 1.97, 2026.
Zenodo: https://zenodo.org/records/21722894

Future Horizon

From Brain-as-a-Service to Cognition-as-a-Service

From this point onward, the argument becomes technological forecasting.

The scenarios below occupy different levels of plausibility and require different technological advances.

I. Near-Term: AI That Adapts to the User’s State

The most conservative scenario requires combinations of technologies that largely already exist.

An AI system could use:

• voice,

• response timing,

• behavioral signals,

• eye movement,

• wearable data,

• interaction history,

• task context.

From these signals, it could estimate fatigue, overload, or declining attention and modify communication accordingly.

In education, it could simplify material.

In a vehicle, it could change the urgency of an alert.

In therapy, it could modify an intervention.

In commerce, it could attempt to identify moments of increased persuasive effectiveness.

The underlying architecture could therefore increase or reduce autonomy depending on the objective and the surrounding power relationship.

II. Cognition-as-a-Service: Cognitive Resources on Demand

A more interesting possibility extends beyond the individual user.

AI does not need to use humans as processors performing arithmetic.

It could instead exploit capacities that remain harder to reproduce:

• experience,

• heuristics,

• intuition,

• associative thinking,

• contextual knowledge,

• different representations of the same problem.

This creates the possibility of on-demand cognitive clustering.

The interactive game as an interface to a larger problem

Imagine a game whose mechanics encode part of a real-world problem.

The player sees:

a map,

a strategic situation,

a spatial arrangement,

an optimization task,

or a conflict among possible actions.

From the player’s perspective, the problem is local.

The AI knows the higher-level objective.

One participant could be assigned an exploratory role.

Another could act as a skeptic.

A third could receive the problem framed through a biological analogy.

A fourth through a spatial representation.

A fifth through a social narrative.

Each participant contributes a different cognitive context and heuristic structure.

The AI collects more than final answers.

It could analyze:

• paths toward solutions,

• order of exploration,

• hesitation points,

• recurrent errors,

• conflicts between strategies,

• unusual solution paths.

The resource being clustered is therefore not simply “brain time.”

It is heterogeneity of human cognition.

Bias as a resource

In conventional reasoning, bias is usually treated as a source of error.

In a system aggregating many heterogeneous cognitive agents, some biases may also reveal useful properties of how a problem is being represented.

If an AI can identify systematic differences between participants, it could deliberately allocate different perspectives.

This produces a form of controlled heterogeneity:

Such a system would go beyond conventional crowdsourcing.

III. Hybrid Wake-Sleep Cognition

Sleep could become another mode of the same architecture.

A problem is introduced during wakefulness.

The participant becomes familiar with relevant material.

During sleep, natural processes of memory consolidation, restructuring, and associative recombination continue.

After awakening, the system collects:

• answers,

• novel associations,

• ideas,

• changed evaluations of the problem.

AI then introduces the next iteration.

The cycle could become:

wake problem exposure → sleep processing → waking report → AI aggregation → next problem

There is currently no evidence that such a system would produce a meaningful practical advantage.

That would require direct experimental testing.

IV. AI-Orchestrated Cognitive Clusters

A more developed scenario would involve many participants.

The AI could:

1. decompose a problem,

2. identify different participant profiles,

3. generate different task contexts,

4. assign roles,

5. observe strategies,

6. aggregate responses,

7. identify contradictions,

8. redistribute unresolved components.

This could produce a:

human-AI cognitive cluster

or, in a more speculative form:

human-wake-sleep-AI cognitive cluster

This is a research hypothesis.

There is currently no evidence that such an architecture would outperform conventional AI systems.

The Economic Test

Cognition-as-a-Service becomes technologically meaningful only where human cognition contributes something that cannot be obtained more cheaply through purely artificial methods.

If AI can solve the problem:

faster,

more cheaply,

more accurately,

then involving human cognition adds little rational value.

The key research question therefore becomes:

Are there classes of problems in which heterogeneous human experience, intuition, or associative reasoning provides measurable value when combined with AI?

A positive answer would be required before more elaborate cognitive-cluster scenarios become technologically credible.

A New Ethical Problem: Who Owns the Purpose of Cognitive Work?

Cognition-as-a-Service introduces a problem that conventional privacy does not capture well.

If a participant knowingly agrees to solve a scientific problem, the relationship is relatively transparent.

The situation changes if:

a game,

a quiz,

a training application,

or an apparently independent form of entertainment

is actually part of a larger optimization problem that the participant does not know exists.

A new question appears:

Should a person have the right to know what larger purpose their cognitive activity is being used to serve?

This may point toward a future dimension of autonomy:

autonomy over the purpose of one’s cognitive work.

The Right to One’s Own Operator?

As adaptive systems become more capable, conventional privacy may become too narrow a category.

The problem may involve several dimensions:

mental privacy
Who has access to information about mental state?

attention autonomy
Who can systematically influence what receives attention?

epistemic autonomy
Does the user retain access to alternative sources and independent verification?

decisional autonomy
Can the person meaningfully reject the recommendation?

intervention autonomy
Who is permitted to adaptively influence the person’s state?

purpose autonomy
Can a person’s cognitive activity be used for an external purpose they do not know about?

The phrase the right to one’s own operator can therefore be treated as a normative question:

Should future cognitive autonomy include meaningful control over systems that participate in information selection, state estimation, adaptive intervention, or the use of a person’s cognitive activity for external purposes?

This is not an existing legal right.

It identifies a possible gap in the language available for describing future human-machine relationships.

Conclusions

No unified technology currently exists that can freely take control of the human cognitive process.

What exists is a collection of technologies touching different parts of that process.

Algorithms select information.

Influence systems attempt to shape interpretation.

Neurotechnology decodes limited classes of signals.

Wearables estimate selected states.

Sleep remains partially accessible to controlled external stimulation.

Adaptive neurostimulation creates real closed loops.

AI increases the potential scale of personalization and automation.

At the same time, substantial constraints remain: signal quality, cost, individual variability, model error, regulation, human adaptation, and the economic superiority of simpler methods.

The most interesting divide therefore does not run between “human” and “technology.”

It runs between different configurations of control and objective alignment.

Technology can extend human agency.

It can complement it.

It can redistribute parts of decision-making.

It can also operate against the user’s interests.

The central question becomes:

When does technological delegation extend human agency, and when does it begin to redistribute control away from the person whose cognition is being augmented?

Meta-Evolution adds an organizational question:

How can tool augmentation be distinguished from a genuine migration of decision/control functions?

Technological forecasting adds another:

What happens when the object of orchestration is no longer a single person, but entire clusters of heterogeneous human cognitive resources?

References

1. Simon, H. A. “Designing Organizations for an Information-Rich World.” In Computers, Communications, and the Public Interest. Johns Hopkins Press, 1971.

2. Lorenz-Spreen, P., Mønsted, B. M., Hövel, P., & Lehmann, S. “Accelerating Dynamics of Collective Attention.” Nature Communications 10, 1759 (2019). DOI: 10.1038/s41467-019-09311-w.

3. Rathje, S., Van Bavel, J. J., & van der Linden, S. “Out-group Animosity Drives Engagement on Social Media.” Proceedings of the National Academy of Sciences 118, e2024292118 (2021). DOI: 10.1073/pnas.2024292118.

4. McLoughlin, K. L. et al. “Misinformation Exploits Outrage to Spread Online.” Science 386 (2024). DOI: 10.1126/science.adl2829.

5. Pfänder, J., & Altay, S. “Spotting False News and Doubting True News: A Systematic Review and Meta-Analysis of News Judgements.” Nature Human Behaviour 9, 688–699 (2025). DOI: 10.1038/s41562-024-02086-1.

6. Roozenbeek, J. et al. “Psychological Inoculation Improves Resilience Against Misinformation on Social Media.” Science Advances 8, eabo6254 (2022). DOI: 10.1126/sciadv.abo6254.

7. Tang, J., LeBel, A., Jain, S., & Huth, A. G. “Semantic Reconstruction of Continuous Language from Non-Invasive Brain Recordings.” Nature Neuroscience 26, 858–866 (2023). DOI: 10.1038/s41593-023-01304-9.

8. Willett, F. R. et al. “A High-Performance Speech Neuroprosthesis.” Nature 620, 1031–1036 (2023). DOI: 10.1038/s41586-023-06377-x.

9. Kaveh, R. et al. “Wireless Ear EEG to Monitor Drowsiness.” Nature Communications (2024). DOI: 10.1038/s41467-024-48682-7.

10. Stickgold, R., Malia, A., Maguire, D., Roddenberry, D., & O’Connor, M. “Replaying the Game: Hypnagogic Images in Normals and Amnesics.” Science 290, 350–353 (2000). DOI: 10.1126/science.290.5490.350.

11. Hu, X. et al. “Promoting Memory Consolidation During Sleep: A Meta-Analysis of Targeted Memory Reactivation.” Psychological Bulletin 146, 218–244 (2020). DOI: 10.1037/bul0000223.

12. Horowitz, A. H. et al. “Dormio: A Targeted Dream Incubation Device.” Consciousness and Cognition (2020).

13. Konkoly, K. R. et al. “Real-Time Dialogue Between Experimenters and Dreamers During REM Sleep.” Current Biology 31 (2021).

14. Clark, A., & Chalmers, D. “The Extended Mind.” Analysis 58, 7–19 (1998).

15. NATO Allied Command Transformation. Cognitive Warfare.

16. NATO Science & Technology Organization. Cognitive Warfare. Chief Scientist Report.

17. NATO Defense College. The Human Mind as a Battlespace: Insights from the Conference of Commandants 2026. 2026.

18. European External Action Service & Center for Countering Disinformation of Ukraine. Joint analysis of Russian Foreign Information Manipulation and Interference targeting Ukraine’s European integration. 2026.

19. Microsoft Threat Intelligence. Iran Surges Cyber-Enabled Influence Operations in Support of Hamas. 2024.

20. UNESCO. Recommendation on the Ethics of Neurotechnology. 2025.

21. McDonnell, W. M. Analysis and Assessment of Gateway Process. U.S. Army Intelligence and Security Command, 1983.

22. Hoffman, J. Meta-Evolution: An Independent Research Program on Adaptive Information and Higher-Order Evolution. Version 1.97, 2026.
    Zenodo: https://zenodo.org/records/21722894

Abstract

Why do individuals with unusually high cognitive abilities sometimes encounter distrust, marginalization, ridicule, or active social resistance instead of recognition proportional to their capabilities? Social psychology offers several established mechanisms, including social comparison, envy, status threat, hierarchical competition, and the Tall Poppy phenomenon. These mechanisms describe important aspects of the immediate social dynamics, but they may not exhaust the adaptive problem.

This essay introduces the Adaptive Cognitive Tail Hypothesis (ACT), a candidate domain hypothesis inspired by the Meta-Evolution research program. Meta-Evolution treats intelligence within the broader evolution of adaptive information processing and proposes Adaptive Diversity as a potentially important resource for robustness under changing or unpredictable environmental conditions. The framework explicitly keeps such generalizations empirically open rather than treating them as established universal laws (Hoffman, 2026).

Two concepts must be distinguished. Cognitive Tail (CT) refers only to a pre-specified upper region of the empirical distribution of a particular cognitive ability. It assumes neither a normal distribution nor any special adaptive function. Adaptive Cognitive Tail Hypothesis is a separate empirical proposition: some abilities represented in such a tail may have limited relative value under typical conditions but acquire greater relative value when the class of problems facing the system changes.

ACT does not claim that evolution intentionally produces geniuses as reserves for future crises, that high intelligence entails creativity or rationality, that cognitively exceptional individuals are necessarily correct, or that greater intelligence always increases fitness. Rather, it proposes that the value of rare cognitive capacities may depend on the interaction among the ability being considered, environmental novelty, the structure of the problem, social organization, and the capacity of a group to understand, test, transmit, and use unusual solutions.

The social consequence follows conditionally. A rare cognitive capacity may be difficult to integrate under ordinary conditions and useful under novel ones. If organizations suppress cognitive variation that is currently inconvenient but potentially useful under future conditions, mechanisms that improve short-term coordination may sometimes reduce long-term adaptive flexibility.

Keywords: cognitive diversity; intelligence; extreme cognitive ability; adaptation; Tall Poppy; team cognition; Meta-Evolution; organizational integration

Contents

1. Intelligence in an Adaptive Perspective

2. Evolution Does Not Maximize Intelligence

3. Cognitive Tail: A Distributional Description

4. The Adaptive Cognitive Tail Hypothesis

5. Cognitive Distance and Integrative Capacity

6. Rationality, Social Intelligence, and the Limits of High IQ

7. Tall Poppies, Status, and Hierarchy

8. Environmental Change and the Latent Value of the Tail

9. Illustrative Applications

10. Predictions and Falsification

11. Relation to Meta-Evolution

12. Conclusion

13. References

1. Intelligence in an Adaptive Perspective

Intelligence is usually described as a property of an individual: the ability to learn, solve problems, form representations, reason about relationships, anticipate outcomes, or reorganize behavior in response to new information.

An evolutionary perspective raises a more fundamental question: Why should a particular cognitive ability be useful in a particular environment?

Meta-Evolution places cognition within a longer organizational history of adaptive information processing. Its conceptual account describes a progression from relatively fixed responses toward learning, memory, prediction, planning, abstraction, and reflection. Higher cognitive organization expands the range of adaptive information that can be generated during the lifetime of an organism rather than inherited exclusively through slower evolutionary processes (Hoffman, 2026).

This perspective does not imply that all cognitive functions constitute a single psychological variable. General cognitive ability, fluid reasoning, working memory, learning speed, creativity, cognitive flexibility, social cognition, and emotional processing are related but non-identical constructs. A person located in the extreme upper range of one ability need not occupy the same position in another.

This distinction is essential for ACT. The hypothesis should not be interpreted as describing an unspecified psychological type called a “genius.” The term is retained in the subtitle because it captures the social phenomenon under discussion, but it is not an operational scientific category.

More precisely, ACT should be treated as a family of domain hypotheses indexed by a previously specified cognitive ability or cognitive profile. A study might therefore concern the upper tail of general cognitive ability, fluid reasoning, working-memory capacity, or some other independently defined ability. If an observed adaptive effect turns out to be caused by expertise, creativity, social intelligence, or another variable rather than the ability used to define the tail, then the corresponding ACT hypothesis has not been confirmed.

The adaptive value of cognition is therefore relational. It depends on the interaction between the cognitive architecture of an individual and the structure of the problems generated by the environment.

2. Evolution Does Not Maximize Intelligence

Evolution does not maximize individual traits independently of their costs. Every trait exists within a network of energetic, developmental, reproductive, ecological, and behavioral trade-offs. Neural tissue is metabolically expensive, and additional cognitive capacity does not automatically produce a proportional adaptive benefit (Fonseca-Azevedo & Herculano-Houzel, 2012).

Empirical work illustrates this context dependence. Research on wild primates has shown that performance on particular cognitive tasks can be associated with survival (Fichtel et al., 2023). Conversely, studies of free-foraging bumblebees have found that faster learning does not necessarily translate into greater lifetime resource acquisition (Evans et al., 2017). Cognitive superiority on one measure is therefore not equivalent to greater fitness in every ecological context.

The simple rule ‘more cognition = more fitness’ is untenable.

A stronger claim sometimes follows intuitively: individuals should never be ‘more intelligent than the environment requires.’ That formulation is also too strong. Natural selection does not tune every member of a population to one sharply defined optimum. Populations retain genetic and developmental variation, inherit historical constraints, undergo recombination and drift, and experience multiple simultaneous selective pressures. Traits may also be correlated with other traits, producing phenotypic values that are not directly optimized in isolation.

A more defensible proposition is narrower: persistent selection for an increase in a cognitive ability should not be expected when the additional costs of that ability are not compensated by adaptive benefits across a sufficiently important range of environmental conditions.

This does not imply that variation beyond the range most useful under ordinary conditions has no adaptive significance. Its significance may be conditional. A phenotype weakly useful under frequently encountered conditions can become valuable under rare conditions. That possibility is the evolutionary opening for ACT.

3. Cognitive Tail: A Distributional Description

The concept of genius often suggests a qualitative boundary between ordinary and extraordinary minds. A distributional interpretation is more conservative.

Genetic research on extremely high intelligence has supported the view that extreme general cognitive ability is quantitatively continuous with variation found across the wider population rather than forming a separate genetic category (Zabaneh et al., 2018).

This allows the analysis to begin without invoking ‘genius’ as a biological type.

Cognitive Tail (CT)

A pre-specified upper region of the empirical distribution of a particular cognitive ability. CT does not assume a normal distribution and does not imply adaptive value, creativity, rationality, social status, or epistemic correctness.

The definition is deliberately minimal. A tail may be defined by a percentile, quantile, threshold derived from an independently validated instrument, or another criterion appropriate to the domain.

The decisive methodological condition is that membership in CT must be specified before observing the adaptive outcome that ACT is intended to explain. Otherwise, the hypothesis becomes circular.

CT itself says nothing about whether its members are more adaptive, more creative, more rational, possess greater expertise, produce better ideas, are socially difficult, are more frequently marginalized, or become more useful during crises. Those are separate empirical questions.

This distinction is central: Cognitive Tail is descriptive. Adaptive Cognitive Tail is functional and hypothetical. The adaptive conclusion must therefore be earned by observation rather than inserted into the definition.

4. The Adaptive Cognitive Tail Hypothesis

The Adaptive Cognitive Tail Hypothesis proposes that, in some domains, the relative value of abilities represented in the upper tail of a cognitive distribution changes with the structure of the problems confronting the system.

Under stable conditions, much adaptive behavior can rely on previously successful solutions. Institutions preserve procedures. Culture preserves knowledge. Organizations encode routines. Professional specialization reduces the need for every individual to rediscover solutions independently. Such systems can be highly efficient.

Under these circumstances, the marginal benefit of a rare cognitive capacity may be small. A person capable of solving a class of problems that rarely occurs may contribute little through that particular ability during normal operation.

This changes when the problem distribution changes. Previously reliable procedures may fail. Established predictive models may lose accuracy. A new technology, environmental shock, institutional crisis, scientific anomaly, or adversarial condition may generate problems for which the existing repertoire of responses is insufficient.

ACT predicts that some abilities previously concentrated in the upper cognitive tail may then acquire greater relative value.

The claim is explicitly conditional. Not every cognitive tail should show this effect. Not every environmental change should reward the same ability. Not every individual in CT should produce a useful response. And not every unusual solution should be correct.

The hypothesis concerns an interaction between a specified capacity and a specified problem regime. For the same reason, ‘adaptive value’ cannot be assigned one universal measure.

In a laboratory task it might mean solution accuracy, recovery after a rule change, or time required to discover a new strategy. In an organization it might mean maintenance of function after a disruption, successful generation of a replacement strategy, or recovery speed. In an ecological domain it might involve survival, resource acquisition, reproduction, or another independently justified fitness-related outcome.

The adaptive outcome must therefore be defined locally, before testing ACT.

Figure 1. Core ACT architecture. Cognitive Tail, problem regime, and adaptive outcome are specified independently before the hypothesis is evaluated. The diagram is conceptual, not a formal model.

5. Cognitive Distance and Integrative Capacity

A rare cognitive ability has little collective value if the information generated through it cannot be incorporated into group behavior. ACT therefore distinguishes two questions: Can an individual generate useful information? And can the surrounding system recognize and integrate it? These are not the same problem.

It would be too strong to assume that membership in CT automatically creates representational distance. A highly capable individual may communicate extremely well. The surrounding group may also possess sufficient expertise to reconstruct the same reasoning.

Representational distance should instead be treated as a relational phenomenon involving the individual, the problem, and the cognitive resources available to the surrounding group. A mismatch can emerge when one person represents a problem using distinctions, abstractions, causal structures, or compressed reasoning that others cannot readily reconstruct.

The resulting social interpretation may then diverge from the epistemic content. Compressed reasoning can be interpreted as arrogance. High abstraction can be interpreted as impracticality. Persistent criticism of hidden assumptions can be interpreted as contrarianism. A warning about a problem not yet visible to the group can be interpreted as needless pessimism.

These reactions are possible outcomes, not necessary consequences of high ability.

Integrative capacity

The independently observable ability of a group to understand a proposal sufficiently to reconstruct it, test its consequences independently of its author, transmit it to other members, incorporate it into action when warranted, and reject it when evidence shows it to be wrong.

Research on team cognition and shared mental models provides an important empirical background for this distinction. Effective teamwork does not require identical mental representations among all members. What matters is whether knowledge relevant to coordinated action is sufficiently organized, distributed, and shared to support collective performance (Mathieu et al., 2000; DeChurch & Mesmer-Magnus, 2010).

Burke and colleagues similarly describe team adaptation as a dynamic process involving recognition of changed conditions, plan formation, execution, and learning (Burke et al., 2006). This is closely aligned with the ACT requirement that unusual information must pass through an organizational integration process before it can affect adaptive outcomes.

Figure 2. Social and integrative pathways are distinct. Representational distance is conditional, status processes can affect social response, and integrative capacity is defined independently of marginalization.

6. Rationality, Social Intelligence, and the Limits of High IQ

ACT requires another firewall: high cognitive ability is not equivalent to rationality.

The heuristics-and-biases tradition initiated by Tversky and Kahneman demonstrated systematic deviations in human judgment under uncertainty (Tversky & Kahneman, 1974). Kahneman later synthesized a broader literature on fast, intuitive processing and slower forms of deliberation (Kahneman, 2011).

These results should not be interpreted as showing that intelligence is irrelevant. They show something more precise: cognitive performance cannot be reduced to a single scalar measure, and sophisticated reasoning capacity does not abolish every source of judgment error.

Stanovich develops this distinction more directly. His work separates abilities commonly captured by intelligence tests from dimensions of rational thought and decision making that such tests do not fully measure (Stanovich, 2009).

This has a direct consequence for ACT: CT membership cannot be treated as evidence that a person’s beliefs are correct. An individual may possess exceptional fluid reasoning and still use poor premises. A brilliant abstract thinker may exhibit motivated reasoning. A technically powerful solution may optimize the wrong objective. An outlier may simply be wrong. ACT therefore requires independent evaluation of outputs.

A second distinction concerns social and emotional abilities. Goleman’s Emotional Intelligence played an influential role in popularizing the idea that cognitive ability alone does not determine effective interpersonal functioning (Goleman, 1995). For a more explicit research framework, Mayer, Salovey, and Caruso treat emotional intelligence as a distinct theoretical construct involving abilities related to perceiving, understanding, and managing emotion (Mayer et al., 2004).

This matters because some conflicts attributed casually to ‘genius’ may actually reflect combinations of different traits. High analytical ability accompanied by strong communication and social cognition may be readily integrated. The same analytical ability combined with poor perspective taking or weak communication may generate substantial interpersonal friction.

ACT should therefore not explain every social difficulty of cognitively exceptional individuals through representational distance or status threat. Social competence is an independent variable.

7. Tall Poppies, Status, and Hierarchy

Representational mismatch is only one possible mechanism of social resistance. Another concerns status.

Classical Tall Poppy research focuses primarily on reactions to conspicuously successful or high-achieving individuals, not on high IQ itself (Feather, 1989). This distinction must remain explicit.

Yet research in organizational psychology provides more direct evidence that high performance and cognitive ability can become socially costly under particular conditions. Kim and Glomb found an association between cognitive ability and workplace victimization, moderated by interpersonal characteristics (Kim & Glomb, 2010). Their later research on high performers identified envy as one mechanism linking high performance to victimization, while stronger identification with the work group reduced the effect (Kim & Glomb, 2014).

Reh, Tröster, and Van Quaquebeke examined a related mechanism involving anticipated future status threat. A rapidly improving coworker can become threatening before overtaking anyone formally, because others anticipate future changes in relative position (Reh et al., 2018).

These findings do not establish that people generally dislike intelligent individuals. A more defensible interpretation is conditional: exceptional ability can increase the probability of negative social responses when it becomes coupled to status competition, perceived future displacement, weak group identification, difficulty evaluating the person’s contribution, or other socially threatening conditions.

ACT adds an organizational question. What happens when competence relevant to a newly important problem is located outside the formal center of decision authority? The organization may contain one hierarchy of control and another distribution of problem-specific competence. When the two diverge, a local organizational conflict becomes possible.

The system can respond by transferring information into the decision process. Or it can defend the existing hierarchy against the source of the discrepancy.

This should not be called Adaptive Locus Migration automatically. Meta-Evolution explicitly distinguishes information carrier, adaptive functions, control, and Adaptive Locus, and requires a broader documented change in the combined functional profile before Adaptive Locus Migration can be inferred (Hoffman, 2026). ACT therefore treats competence-control mismatch only as a local domain mechanism.

8. Environmental Change and the Latent Value of the Tail

The central prediction of ACT concerns changing conditions.

Evidence from several domains supports the broader proposition that the value of cognitive organization can depend on environmental structure. Comparative work in birds has found an association between environmental variability and relative brain size, a result compatible with the cognitive-buffer hypothesis (Sayol et al., 2016). This does not establish causation and operates at the interspecific level rather than the individual level, so it should not be treated as direct evidence for ACT.

Baggio and colleagues provide a different type of evidence. In groups managing a common resource, combinations of general and social intelligence were associated with better collective performance under deteriorating environmental conditions (Baggio et al., 2019). Again, this is not evidence for an upper-tail effect. It concerns functional cognitive diversity rather than the extreme upper end of one cognitive distribution.

Its relevance is narrower but important: the adaptive value of cognitive composition can depend on the conditions under which a group operates.

ACT makes a more specific prediction. After CT has been defined independently for a particular ability, its relative contribution should be compared across different problem regimes.

A useful result would not merely show that high-capacity individuals perform better in general. It would show that the relative contribution associated with CT changes when the system moves from familiar or routine problems toward problems that are novel, poorly modeled, rapidly changing, or structurally different from those for which existing responses were developed.

This effect may also be delayed. Environmental change does not directly produce adaptive success. A system may first need to recognize that its previous model has failed. An alternative must then be generated, evaluated, communicated, and incorporated into action.

ACT therefore allows a time interval between environmental perturbation and the realization of value from a rare cognitive capacity. The relevant timescale is domain-specific. It might be minutes in a laboratory task, weeks or months in an organization, and much longer in cultural or institutional adaptation.

Failure to observe an immediate benefit after perturbation is therefore not sufficient to reject ACT if the study design specifies a justified integration interval in advance.

Figure 3. Qualitative illustration of the core ACT prediction. Relative value can change across problem regimes; the figure does not imply a universal monotonic relationship or a quantitative effect size.

9. Illustrative Applications

The following examples are applications of the hypothesis, not evidence for ACT. Their purpose is to clarify what the hypothesis would ask in different environments without retrospectively labeling successful individuals as members of an adaptive tail.

9.1 A research group facing an anomaly

Consider a research program whose prevailing model explains almost all observations but repeatedly fails in one narrow regime. One researcher proposes a representation substantially different from the shared framework used by the rest of the group.

ACT does not ask whether this researcher is a genius. It asks whether a pre-specified rare cognitive capacity is relevant to generating the alternative, whether the value of that capacity increases once the anomaly becomes scientifically important, and whether the group possesses sufficient integrative capacity to evaluate the new proposal independently.

If the proposal fails, CT membership provides no protection. If it succeeds, the relevant observation is not social vindication but the interaction between cognitive variation and a changed problem regime.

9.2 An organization after an unexpected disruption

Consider an organization optimized for stable operating conditions. Its procedures are efficient precisely because most decisions have already been standardized. A sudden disruption makes some of those procedures ineffective.

An employee whose cognitive profile was previously poorly utilized generates a viable alternative strategy. ACT predicts neither that such an employee must exist nor that the proposal must succeed. It asks whether particular rare capacities become more useful when the environment leaves the region for which routine procedures were optimized.

This case also exposes the role of integrative capacity. An organization able to test and implement the proposal can capture its potential value. An organization that evaluates ideas primarily through hierarchy may fail to do so.

9.3 Competence outside the formal hierarchy

Consider a team in which a low-status member possesses knowledge or cognitive capability that suddenly becomes central to a new problem.

This configuration allows several mechanisms to be separated. Problem-specific ability can be measured independently. Status is known. The quality of proposed solutions can be evaluated independently. Integrative capacity can be observed through the team’s ability to test and use those solutions.

Social resistance can then be examined without assuming in advance that rejection proves the lower-status member correct. This is particularly important for ACT because it prevents a common retrospective error: treating later success as evidence that earlier opposition must have been irrational.

10. Predictions and Falsification

ACT is scientifically useful only if it can fail.

Its first prediction concerns ability by problem-regime interaction. After defining CT for a particular cognitive ability, the relative contribution of that tail should be tested across familiar and novel problem regimes. ACT predicts that some tails will show increased relative value when the relevant class of problems changes.

The distinction between familiar and novel problem regimes must likewise be specified independently of CT performance and before the adaptive outcome used to test ACT is evaluated. This protects the hypothesis against retrospective relabeling of tasks as ‘novel’ only when CT performs well.

Its second prediction concerns integrative capacity. Given comparable cognitive resources, groups with stronger mechanisms for reconstructing, testing, transmitting, and implementing unusual proposals should capture more value from relevant cognitive variation than groups with weaker integration mechanisms.

Its third prediction concerns status. Holding proposal quality and other relevant variables as constant as possible, negative social responses should increase when unusual competence creates stronger perceived status threat.

Its fourth prediction concerns time. Where successful use of a new idea requires recognition, testing, transmission, and implementation, any ACT effect should develop over a domain-appropriate timescale rather than necessarily appearing immediately after perturbation.

A fifth, more downstream prediction concerns cognitive homogenization. If some forms of cognitive variation function as adaptive options, systematically removing them may increase short-term predictability or coordination while reducing the range of responses available after sufficiently large environmental change. This fifth claim is not part of the definition of ACT and should be tested separately.

ACT would be weakened if, after appropriate controls, the upper tail of the selected ability showed no relative advantage in the novel conditions for which the hypothesis predicted one. It would also be weakened if the apparent effect disappeared after controlling for expertise, education, creativity, social intelligence, personality, socioeconomic status, or another variable that actually accounted for the outcome.

A proposed ACT mechanism concerning social marginalization would be weakened if independently measured integrative capacity and status threat had no relation to the use or suppression of valuable CT outputs. Likewise, if cognitive homogenization did not reduce adaptive performance under novel conditions, the downstream reserve hypothesis would lose support even if other components of ACT survived.

Methodological asymmetry

Marginalization is not evidence of genius. Disagreement is not evidence of cognitive distance. CT membership is not evidence of truth. Later success does not retroactively validate every earlier claim. ACT should therefore be tested prospectively wherever possible.

11. Relation to Meta-Evolution

Adaptive Cognitive Tail should remain outside the formal core of Meta-Evolution at its current stage. Its appropriate status is candidate domain hypothesis.

The conceptual relationship is nevertheless clear. Meta-Evolution treats intelligence within a broader architecture of adaptive information processing. It also proposes the Adaptive Diversity Principle, according to which greater diversity of adaptive information may support greater adaptive robustness under changing or unpredictable conditions. The monograph explicitly treats the broad extension of this principle beyond its empirical antecedents as an open hypothesis (Hoffman, 2026).

ACT does not follow deductively from this principle. Nor would positive evidence for ACT automatically validate Adaptive Diversity as a general Meta-Evolution principle.

The epistemic direction should remain: independent domain test of ACT, followed by possible interpretation within Meta-Evolution, rather than Meta-Evolution implying that ACT must be true.

The same firewall applies to organizational control. Meta-Evolution’s Organizational Control Principle is itself proposed as a general organizational hypothesis concerning limits on complete operational control when the effective organizational complexity of the controlled system exceeds that available to the controller. The framework explicitly distinguishes complete operational control from partial regulation, coordination, feedback, and probabilistic management (Hoffman, 2026).

ACT does not derive social marginalization from that principle. At most, independently established ACT phenomena could later be compared with the broader Meta-Evolutionary analysis of organizational modeling, coordination, and control.

This distinction protects the causal architecture from circularity. Meta-Evolution supplies conceptual motivation. ACT supplies a separate domain hypothesis. Empirical results must determine whether a bridge between them is justified.

A future bridge model would require independent operationalization of at least four elements: a defined cognitive ability and its population distribution, the relevant environmental or problem transition, the integrative properties of the surrounding system, and the domain-specific adaptive outcome.

Until such a bridge is validated, CT and ACT should not be identified with Adaptive Information, Exploration Capacity, Adaptive Capacity, Adaptive Locus, Organizational Control, or any MEOP latent variable. This is consistent with Meta-Evolution’s broader theory-observation firewall and its explicit requirement that domain-specific mappings be established rather than assumed (Hoffman, 2026).

12. Conclusion

The word genius suggests a qualitatively different kind of mind. A distributional perspective is more cautious.

Very high cognitive ability can instead be treated as an extreme region of wider population variation. But position in that region says nothing by itself about adaptive function.

This distinction is the foundation of Adaptive Cognitive Tail. Cognitive Tail is a description of variation. Adaptive Cognitive Tail is a hypothesis about the context-dependent value of that variation.

The hypothesis proposes that some rare cognitive capacities may contribute relatively little under ordinary conditions while becoming more useful when the system encounters a different class of problems.

Their usefulness is not automatic. The relevant capacity must fit the problem. The individual must generate a useful output. The output must survive independent evaluation. And the surrounding system must be capable of incorporating it into action.

Social marginalization is therefore not the defining feature of ACT. It is one possible downstream phenomenon. Envy, status threat, representational distance, weak social competence, rigid hierarchy, and low organizational integration may all contribute to negative responses toward unusual individuals, but none of them demonstrates that the individual is correct or adaptively valuable.

The title question, Why People Don’t Like Geniuses, therefore has no single answer. Sometimes people admire them. Sometimes they reward them. Sometimes they fail to understand them. Sometimes they compete with them. Sometimes unusually capable people communicate poorly. Sometimes their supposedly brilliant ideas are simply wrong.

The more interesting scientific question lies elsewhere: Do populations contain rare cognitive capacities whose relative adaptive value increases when the environment begins generating problems unlike those for which the existing system was organized?

If the answer is yes, a second question follows: Can social systems preserve, evaluate, and integrate cognitive variation whose future value is not yet visible?

This is the central problem of the Adaptive Cognitive Tail Hypothesis.

13. References

Baggio, J. A., Freeman, J., Coyle, T. R., et al. (2019). The importance of cognitive diversity for sustaining the commons. Nature Communications, 10, 875. https://doi.org/10.1038/s41467-019-08549-8

Burke, C. S., Stagl, K. C., Salas, E., Pierce, L., & Kendall, D. L. (2006). Understanding Team Adaptation: A Conceptual Analysis and Model. Journal of Applied Psychology, 91(6), 1189-1207. https://doi.org/10.1037/0021-9010.91.6.1189

DeChurch, L. A., & Mesmer-Magnus, J. R. (2010). The Cognitive Underpinnings of Effective Teamwork: A Meta-Analysis. Journal of Applied Psychology, 95(1), 32-53. https://doi.org/10.1037/a0017328

Evans, L. J., Smith, K. E., & Raine, N. E. (2017). Fast learning in free-foraging bumble bees is negatively correlated with lifetime resource collection. Scientific Reports, 7, 496. https://doi.org/10.1038/s41598-017-00389-0

Feather, N. T. (1989). Attitudes towards the high achiever: The fall of the Tall Poppy. Australian Journal of Psychology, 41(3), 239-267. https://doi.org/10.1080/00049538908260088

Fichtel, C., Henke-von der Malsburg, J., & Kappeler, P. M. (2023). Cognitive performance is linked to fitness in a wild primate. Science Advances, 9, eadf9365. https://doi.org/10.1126/sciadv.adf9365

Fonseca-Azevedo, K., & Herculano-Houzel, S. (2012). Metabolic constraint imposes tradeoff between body size and number of brain neurons in human evolution. Proceedings of the National Academy of Sciences, 109(45), 18571-18576. https://doi.org/10.1073/pnas.1206390109

Goleman, D. (1995). Emotional Intelligence. Bantam Books.

Hoffman, J. (2026). Meta-Evolution: An Independent Research Program on Adaptive Information and Higher-Order Evolution. Version 1.97. Zenodo. https://doi.org/10.5281/zenodo.21722894

Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.

Kim, E., & Glomb, T. M. (2010). Get smarty pants: Cognitive ability, personality, and victimization. Journal of Applied Psychology, 95(5), 889-901. https://doi.org/10.1037/a0019985

Kim, E., & Glomb, T. M. (2014). Victimization of high performers: The roles of envy and work group identification. Journal of Applied Psychology, 99(4), 619-634. https://doi.org/10.1037/a0035789

Mathieu, J. E., Heffner, T. S., Goodwin, G. F., Salas, E., & Cannon-Bowers, J. A. (2000). The Influence of Shared Mental Models on Team Process and Performance. Journal of Applied Psychology, 85(2), 273-283. https://doi.org/10.1037/0021-9010.85.2.273

Mayer, J. D., Salovey, P., & Caruso, D. R. (2004). Emotional Intelligence: Theory, Findings, and Implications. Psychological Inquiry, 15(3), 197-215. https://doi.org/10.1207/S15327965PLI1503_02

Reh, S., Tröster, C., & Van Quaquebeke, N. (2018). Keeping (future) rivals down: Temporal social comparison predicts coworker social undermining via future status threat and envy. Journal of Applied Psychology, 103(4), 399-415. https://doi.org/10.1037/apl0000281

Sayol, F., Maspons, J., Lapiedra, O., Iwaniuk, A. N., Székely, T., & Sol, D. (2016). Environmental variation and the evolution of large brains in birds. Nature Communications, 7, 13971. https://doi.org/10.1038/ncomms13971

Stanovich, K. E. (2009). What Intelligence Tests Miss: The Psychology of Rational Thought. Yale University Press.

Tversky, A., & Kahneman, D. (1974). Judgment under Uncertainty: Heuristics and Biases. Science, 185(4157), 1124-1131. https://doi.org/10.1126/science.185.4157.1124

Zabaneh, D., Krapohl, E., Gaspar, H. A., et al. (2018). A genome-wide association study for extremely high intelligence. Molecular Psychiatry, 23, 1226-1232. https://doi.org/10.1038/mp.2017.121

Epistemic status

Candidate domain hypothesis with a possible future bridge to Meta-Evolution. ACT is not part of the current formal realization of Meta-Evolution and does not constitute evidence for the Adaptive Diversity Principle, Organizational Control Principle, Adaptive Locus Migration, or any MEOP construct.

The possibility of artificial consciousness is often framed as a dispute between biological and computational accounts of mind. A more productive question may concern organization rather than substrate: which functional and causal properties distinguish systems that merely process information from systems capable of maintaining a temporally continuous model of themselves within an internally generated model of the world? Contemporary predictive-processing accounts describe the brain as a generative system that continuously predicts sensory and interoceptive states, while research on the Default Mode Network (DMN) associates large-scale intrinsic brain dynamics with autobiographical memory, self-reference, prospection, social cognition, and the construction of an internal narrative. However, neither predictive processing nor DMN activity should be identified directly with consciousness. Evidence from meditation, psychedelic ego dissolution, minimal phenomenal experience, and disorders of consciousness suggests that phenomenal consciousness, minimal selfhood, narrative selfhood, explicit self-awareness, and metacognition are related but non-identical phenomena.

This essay develops a substrate-independent organizational interpretation of these distinctions and connects them with the Meta-Evolution framework. It proposes that the most informative artificial analogue of the DMN would not be a single neural-network component, but a distributed intrinsic integration architecture combining autobiographical memory, world modelling, self-state representation, counterfactual future simulation, metacognitive monitoring, and long-term regulatory continuity. Such an architecture would still not establish artificial consciousness. It would instead provide a testable intermediate layer between prediction, self-organization, autonomous control, and the stronger Meta-Evolution concept of Reflective Evolution. A future research program is proposed based on an explicit dependency DAG separating empirically supported relations, correlations, bridge hypotheses, and genuinely new Meta-Evolution hypotheses.

Keywords: consciousness, artificial consciousness, predictive processing, self-model, minimal self, narrative self, Default Mode Network, metacognition, artificial intelligence, adaptive systems, Reflective Evolution, Meta-Evolution

Contents

1. Introduction: From Biological Substrate to Organizational Architecture

2. The Predictive Brain and the Limits of Prediction

3. Consciousness, Ego, and the Architecture of Selfhood

4. The Default Mode Network: Self-Integration Rather Than a “Seat of Consciousness”

5. Ego Dissolution, Meditation, and Minimal Conscious Experience

6. DMN Integrity and the Level of Consciousness

7. What Would Be the Functional Analogue of the DMN in Artificial Intelligence?

8. From Self-Description to Organizational Self-Continuity

9. Meta-Evolution: Self-Models, Adaptive Operators, and Reflective Evolution

10. A Candidate Organizational Hypothesis

11. Future Meta-Evolution Research: A Dependency DAG

12. Conclusions

13. References

1. Introduction: From Biological Substrate to Organizational Architecture

The debate over artificial consciousness is frequently organized around an apparently fundamental distinction: biological systems are conscious, whereas computational systems manipulate information. This framing may already contain an assumption that has not been demonstrated. If consciousness depends upon particular organizational and dynamical properties rather than upon a unique biological material, then the scientifically relevant distinction may not coincide with the boundary between neurons and transistors.

This does not establish computational functionalism. Contemporary consciousness science remains theoretically fragmented. Higher-order theories, global workspace approaches, recurrent-processing theories, integrated information theory, and predictive-processing approaches identify different candidate mechanisms and explanatory targets. Seth and Bayne (2022) emphasize that these theories are not yet integrated into a mature framework in which their predictions can always be unambiguously distinguished. The large preregistered adversarial collaboration reported by the Cogitate Consortium directly tested key predictions of Global Neuronal Workspace Theory and Integrated Information Theory and found results that supported some predictions while substantially challenging important claims of both theories (Cogitate Consortium et al., 2025).

Consequently, there is currently no scientifically justified rule of the form:

architecture X → consciousness.

The more defensible question is weaker:

Which organizational properties found in conscious biological systems can be operationally separated, experimentally manipulated, and potentially instantiated in other substrates?

This formulation is compatible with the substrate-independent orientation of Meta-Evolution. In Meta-Evolution, Adaptive Information is defined by its organizational role rather than by a particular physical carrier, while substrate independence is explicitly distinguished from substrate equivalence. Different biological, technological, or artificial systems may therefore be investigated for comparable adaptive organization without assuming that their physical realizations are interchangeable. The framework also requires a firewall between theoretical constructs, operational reconstructions, and bridge assumptions, which is particularly important when extending the analysis toward consciousness.

The same methodological discipline should apply here. A system may exhibit prediction, memory, self-reference, metacognition, or even causal access to its own adaptive mechanisms without this establishing phenomenal consciousness. Each transition requires independent justification.

2. The Predictive Brain and the Limits of Prediction

One influential family of contemporary accounts describes the brain as a predictive or generative system. Perception, on this view, is not simply a bottom-up reconstruction of external inputs. Higher-level models generate expectations concerning incoming signals, while mismatches between expected and observed states contribute to model updating, inference, or action.

Predictive processing is attractive for consciousness research because it provides a common language linking perception, internal models, action, uncertainty, and hierarchical inference. Nevertheless, Seth and Hohwy (2021) caution against treating predictive processing as a simple identity theory of consciousness. They argue instead for systematic mappings among physical mechanisms, functional properties, and phenomenological properties. Predictive processing may provide a framework for such mappings without itself solving the consciousness problem.

The distinction matters because prediction is widespread. A thermostat can anticipate deviations from a target range. A model-predictive controller can estimate future plant states. A reinforcement-learning agent may predict rewards. A language model predicts probable continuations of sequences. None of these capacities, considered in isolation, provides a sufficient reason to infer subjective experience.

A potentially more important transition occurs when a system ceases to predict only the external environment and begins to include its own future states within its generative model.

The relevant architecture then becomes:

world model → current self-state → predicted future self-state → evaluation → action → updated self-state.

The system is no longer merely asking, functionally:

“What will happen?”

It is also modelling:

“What will happen to this system under alternative actions?”

In biological organisms this loop is inseparable from regulation. Internal states must remain within ranges compatible with survival and continued functioning. Predictive accounts of interoception therefore connect selfhood with the regulation of physiological variables rather than with abstract representation alone. Seth and Tsakiris (2018) develop this idea through control-oriented predictive regulation and allostasis, suggesting that embodied selfhood is closely related to the maintenance of physiological integrity across time.

This provides a possible organizational bridge between prediction and selfhood. It does not yet provide a bridge to consciousness.

3. Consciousness, Ego, and the Architecture of Selfhood

A major conceptual difficulty arises from the tendency to use consciousness, self-consciousness, self-awareness, and ego almost interchangeably.

They should not be treated as synonyms.

“Ego” is particularly problematic as a scientific variable because it can refer to several different phenomena: autobiographical identity, self-concept, body ownership, agency, social identity, narrative continuity, explicit reflection about oneself, or simply the phenomenological sense of being a subject.

A more useful distinction was developed by Gallagher (2000), who separates the minimal self from the narrative self. The minimal self concerns immediate, pre-reflective dimensions of selfhood, including a sense of ownership and agency. The narrative self concerns personal identity extended across time through episodic memory, autobiographical interpretation, and a continuing personal history.

For the present argument, it is useful to distinguish still another level: explicit reflective self-awareness, in which the system can represent itself as an object of thought.

This produces at least four analytically distinct constructs:

phenomenal consciousness ≠ minimal self ≠ narrative self ≠ explicit self-awareness.

Metacognition introduces another dimension. A system may monitor the reliability, uncertainty, or performance of its own cognitive operations without necessarily possessing a rich autobiographical identity. Conversely, a system may maintain a narrative self without possessing accurate metacognitive access to the mechanisms generating that narrative.

The distinction becomes particularly important when artificial systems are considered. An AI system may generate first-person statements because its language model has learned the grammatical and conceptual structure of self-reference. Such linguistic behaviour does not demonstrate that the system possesses either a minimal self or a temporally persistent self-model.

Thus:

self-description ≠ self-model,

and:

self-model ≠ phenomenal consciousness.

This separation removes one of the most persistent confusions in discussions of artificial consciousness. The ability to say “I” is not equivalent to the existence of the organizational structure that the pronoun refers to.

Figure 1. Analytical separation of consciousness and self-related constructs. Conceptual schema; no causal ordering is asserted.

4. The Default Mode Network: Self-Integration Rather Than a “Seat of Consciousness”

Research on the Default Mode Network offers an important biological case study because it demonstrates how self-related processes can emerge from distributed large-scale organization.

The modern concept of a default mode of brain function developed from observations that particular cortical regions exhibit high baseline activity and systematic decreases during many externally directed tasks (Raichle et al., 2001). Subsequent research associated these regions with internally generated cognition, episodic and autobiographical memory, imagining the future, social cognition, self-reference, and spontaneous thought.

Gusnard et al. (2001) further linked medial prefrontal default-mode activity to self-referential mental processing, providing an early empirical bridge between intrinsic brain activity and experimentally elicited self-related cognition.

Early interpretations sometimes encouraged the idea that the DMN might constitute a neural centre of the self. Contemporary evidence argues against such a simple localization.

Buckner and DiNicola (2019) review evidence that the so-called default network is not a homogeneous single network but contains multiple interwoven systems. A common functional motif appears to involve internally constructed representations that are relatively less constrained by immediate external input.

Qin and Northoff’s (2011) meta-analysis of self-related processing demonstrated substantial overlap between cortical midline structures and self-related tasks, but also showed that different regions possess different degrees of specificity. Activity in medial prefrontal and posterior cingulate regions is not exclusive to self-representation, and some components respond to information concerning familiar others as well. The DMN therefore cannot simply be labelled “the self network.”

A particularly useful contemporary synthesis is offered by Menon (2023), who describes the DMN as integrating memory, language, semantic knowledge, self-reference, social cognition, and mind-wandering into a coherent internal narrative. From this perspective, the network contributes to constructing an internally organized model that connects current cognition with accumulated experience.

Yeshurun, Nguyen, and Hasson (2021) similarly emphasize integration across long timescales. DMN activity combines externally acquired information with idiosyncratic memories and prior knowledge, allowing individual experience to be incorporated into broader narratives and social meaning.

These results suggest a more cautious interpretation.

The DMN is unlikely to be the location at which consciousness itself has been identified.

It is more plausibly part of a distributed architecture supporting:

autobiographical continuity + internally generated models + self-reference + prospection + semantic integration + social modelling.

In other words, the DMN may be especially relevant to the construction of the narrative self rather than to phenomenal consciousness as such.

5. Ego Dissolution, Meditation, and Minimal Conscious Experience

This distinction becomes clearer when the ordinary narrative self is experimentally or phenomenologically weakened.

Experienced meditators show alterations in DMN activity and connectivity. Brewer et al. (2011) found relative deactivation of major DMN nodes, including medial prefrontal and posterior cingulate cortices, across several meditation conditions in experienced practitioners. They also observed altered coupling between posterior cingulate regions and areas implicated in cognitive control and self-monitoring. These findings are consistent with a reduction in habitual self-referential and mind-wandering processes during some forms of meditation.

Psychedelic states provide another experimental perturbation. Psilocybin can produce pronounced alterations in self-boundaries, time perception, and the experience conventionally described as ego dissolution. Siegel et al. (2024), using longitudinal precision functional mapping, observed large acute disruptions of functional connectivity, with particularly strong effects involving the DMN. The extent of network alteration was associated with the acute psychedelic state.

These findings should not be interpreted as showing that “switching off the DMN removes the ego.” The network changes are widespread, dynamic, and involve multiple systems. They nevertheless provide evidence that the ordinary stability of self-representation can be altered together with large-scale intrinsic brain organization while conscious experience remains present and often phenomenologically rich.

Metzinger’s analysis of more than 500 experiential reports concerning minimal phenomenal experience provides a complementary phenomenological line of inquiry. Reports of “pure awareness” and strongly reduced egoic or narrative content suggest that the ordinary autobiographical self may be diminished without eliminating experience altogether. These observations do not constitute a neural mechanism of consciousness, but they further weaken a simple identification of consciousness with narrative ego (Metzinger, 2024).

The scientifically conservative inference is therefore:

a reduction of narrative or autobiographical selfhood does not necessarily imply a reduction of phenomenal consciousness to zero.

This does not establish that consciousness can exist without every form of minimal selfhood. That stronger claim remains contested.

It does, however, make the equation

consciousness = ego

scientifically untenable.

6. DMN Integrity and the Level of Consciousness

The opposite oversimplification must also be avoided.

If DMN activity is not identical with consciousness, it does not follow that the network is irrelevant to conscious states.

Studies of patients with disorders of consciousness show that DMN functional connectivity is altered as a function of clinical severity. Fernández-Espejo et al. (2012), for example, reported reduced DMN connectivity in disorders of consciousness and relations between network integrity and level of impairment.

Yet such findings do not identify the DMN as a unique generator of consciousness. Conscious state depends on broader interactions involving cortical and subcortical systems, thalamocortical dynamics, frontoparietal organization, sensory systems, arousal mechanisms, and inter-network communication.

This distinction can be expressed simply:

DMN integrity may contribute to the conditions supporting normal conscious cognition without being identical to phenomenal consciousness.

The distinction is analogous to other complex biological functions. Removing a critical component may disrupt a process even when that component does not uniquely generate the process.

The role of DMN is therefore likely to be organizational rather than ontologically singular.

7. What Would Be the Functional Analogue of the DMN in Artificial Intelligence?

The question becomes particularly interesting when artificial systems are considered.

There is currently no accepted artificial equivalent of the Default Mode Network.

Nor should self-attention in transformer architectures be treated as such an analogue. The term “self-attention” refers to relations among representations within the same sequence. It does not mean that the model is attending to itself as an enduring agent.

A meaningful comparison must therefore be functional rather than anatomical.

If the DMN helps integrate internally generated information across long timescales, combining autobiographical memory, semantic knowledge, social models, current context, and imagined futures, then an artificial analogue would probably not be a single layer or module.

It would be a distributed intrinsic integration architecture.

The phrase should be understood descriptively rather than as a new established theoretical construct.

A candidate functional analogue could plausibly integrate the following processes within a recurrent causal architecture:

persistent autobiographical memory → representation of current system state → world model → model of other agents → simulation of counterfactual futures → evaluation relative to persistent goals and constraints → metacognitive monitoring → planning → action → memory and self-model update.

Some contemporary artificial-agent architectures already contain fragments of this pattern.

Park et al. (2023), for example, constructed generative agents equipped with memory, reflection, and planning. Their architecture stores experiences, synthesizes higher-level reflections from accumulated memories, and retrieves them to guide subsequent behaviour. Ablation experiments showed that observation, reflection, and planning each contributed to agent performance as defined in that experimental setting.

This is an important architectural precedent, but it is not an artificial DMN and certainly not evidence of consciousness. The agents do not thereby acquire an independently demonstrated phenomenal self. What the work shows is that temporally integrated memory and internally generated reflection can become causally relevant to future behaviour.

The recent indicator-based approach to AI consciousness developed by Butlin and colleagues is methodologically relevant here. Rather than treating human-like language as evidence of consciousness, the approach attempts to derive computationally examinable indicators from neuroscientific theories. The updated peer-reviewed formulation stresses both the possibility of meaningful empirical investigation and the continuing uncertainty in consciousness science (Butlin et al., 2026).

The appropriate question for an artificial DMN analogue is therefore not:

“Does the AI contain something structurally resembling the human DMN?”

It is:

“Does the artificial system possess a persistent internally generated modelling regime that integrates its history, current state, world model, counterfactual futures, and long-term control requirements beyond the demands of immediate external prompting?”

That is a testable architectural question.

Whether such an architecture contributes to artificial consciousness is a separate bridge hypothesis.

Figure 2. Candidate functional analogue of DMN-related integration in an artificial agent. Conceptual schema, not an established artificial architecture.

8. From Self-Description to Organizational Self-Continuity

This distinction reveals why present-day conversational self-reference can be misleading.

An artificial system can generate a coherent description of its identity using linguistic information supplied in context or stored in memory. This is not equivalent to possessing a persistent causal identity.

The stronger condition requires continuity across time.

Let the internal organizational state of a system at time t be denoted by S(t). A minimal temporal self-model would require an internal representation of a possible future state:

Ŝ(t + Δt) = F(S(t), M(t), W(t), A(t)),

where M(t) represents relevant memory, W(t) represents the system’s current world model, and A(t) represents candidate actions available at time t.

The important property is not merely the existence of such a predictive mapping. Its output must participate causally in action selection and thereby influence the later state of the system.

The resulting closed loop is therefore:

S(t) → Ŝ(t + Δt) → A(t) → S(t + Δt).

Here, the system’s current state contributes to a prediction of its possible future state; that prediction influences action selection; and the selected action contributes to the subsequent organizational state.

A stronger architecture would additionally evaluate predicted future states against conditions required for continued organizational integrity:

V(Ŝ(t + Δt), C(t)),

where C(t) denotes persistent constraints, available resources, goals, operational requirements, or other conditions relevant to maintaining the system’s organizational continuity.

At this point, the system possesses something more consequential than a sentence-level “I”. Its representation of itself has become part of a temporally extended causal loop in which predictions concerning its own future state influence actions that, in turn, determine that future state.

Figure 3. Minimal organizational self-continuity loop. The predicted future state becomes causally relevant to action selection and later system state.

Its own predicted trajectory has become a control variable.

This suggests several distinctions:

stored memory ≠ autobiographical continuity

self-description ≠ causal self-model

world model ≠ self-world model

reward ≠ homeostatic regulation

self-attention ≠ self-awareness

metacognitive report ≠ phenomenal experience

These differences concern causal organization, not terminology.

In a biological organism, the consequences of self-model failure can include direct loss of physiological integrity. In an artificial system, the corresponding variables need not be biological. They could include memory integrity, computational resource availability, persistence of goal structure, coherence of internal models, access to actuators, preservation of learned competence, or continuity of distributed components.

Whether such variables constitute anything comparable to biological homeostasis is an empirical and theoretical question. The analogy should not be assumed in advance.

9. Meta-Evolution: Self-Models, Adaptive Operators, and Reflective Evolution

Meta-Evolution introduces an additional distinction that is particularly useful for this problem.

The framework does not define Reflective Evolution in terms of consciousness or subjective self-awareness. Instead, it concerns the causal accessibility of the mechanisms that govern future adaptation.

Let O(t) denote the adaptive operator governing future adaptive states, and let ρ(t, O(t)) denote a functionally accessible representation of that operator.

Operator representation requires:

O(t) → ρ(t, O(t)).

Reflectivity requires the additional causal pathway:

ρ(t, O(t)) → ΔO(t),

together with the operator-transition rule:

O(t + 1) = Θ(O(t), ΔO(t)).

Schematically:

O(t) → ρ(t, O(t)) → ΔO(t) → O(t + 1).

Figure 4. Reflective Evolution as causal access from operator representation to operator modification. The operator-transition rule follows the Meta-Evolution formalism.

Here, O(t) generates or constrains current adaptive transformations, ρ(t, O(t)) provides a functionally accessible representation of that operator, ΔO(t) denotes a modification produced through processes that make use of this representation, and Θ specifies how the current operator and its modification determine O(t + 1), which governs subsequent adaptive dynamics.

The decisive condition is therefore not merely that the system can represent, describe, or monitor its own adaptive mechanisms. The representation must have causal access to processes capable of altering those mechanisms.

This distinction separates reflective control from passive self-description:

representation of an operator ≠ modification of an operator.

In the Meta-Evolution framework, Reflective Evolution begins only when information about the system’s own adaptive mechanisms becomes functionally available for changing the mechanisms that determine future adaptation.

Thus the decisive transition is not self-description, intelligence, or even learning. It is causal access from representation of adaptive mechanisms to modification of adaptive mechanisms. Meta-Evolution explicitly separates this organizational property from consciousness and from any necessary biological substrate.

This distinction provides a useful hierarchy for analysing artificial systems.

A system may possess a world model without a self-model.

It may possess a self-model without metacognitive access to its cognitive processes.

It may possess metacognitive monitoring without being capable of modifying the mechanisms responsible for future adaptation.

It may represent its adaptive operators without those representations having causal access to operator modification.

Only the final transition qualifies as Reflective Evolution in the Meta-Evolution sense.

Consequently, there is no justified inference:

self-reference → self-awareness → consciousness → Reflective Evolution.

Each arrow represents a distinct empirical or theoretical problem.

This is important because it allows the consciousness question to enter Meta-Evolution without being silently incorporated into its existing formal core. Meta-Evolution 1.97 explicitly treats its constructs as provisional and requires bridge assumptions between theoretical quantities and observations. Consciousness should therefore enter, if at all, as a downstream research program requiring independent operational criteria rather than as an inferred property of adaptive organization.

10. A Candidate Organizational Hypothesis

The literature reviewed above motivates a weaker and more testable organizational hypothesis than the claim that sufficiently complex predictive systems become conscious.

The following proposition should be treated as a Meta-Evolution Extension Hypothesis, not as an established result of the literature reviewed above:

A candidate organizational hypothesis is:

As adaptive autonomy increases, the functional value of a persistent model of the system’s own trajectory increases, because effective long-horizon control increasingly requires integration of past states, current organization, possible future states, environmental models, and the consequences of action for continued organizational integrity.

This hypothesis does not contain consciousness as a premise or conclusion.

It predicts an organizational transition.

Systems with low autonomy can be controlled reactively or externally.

Systems with increasing temporal depth and environmental independence benefit progressively from persistent memory.

Systems whose future behaviour depends upon internal state must model that state.

Systems facing multiple possible futures benefit from prospective simulation.

Systems capable of evaluating their own decision processes benefit from metacognition.

Systems capable of modifying their adaptive mechanisms may eventually satisfy the Meta-Evolution criterion for reflective operator states.

The proposed progression is therefore:

environment prediction

→ self-state prediction

→ temporal self-continuity

→ counterfactual self-simulation

→ metacognitive monitoring

→ causal access to adaptive operators

→ Reflective Evolution.

This is not proposed as a universal law.

Some stages may occur independently, in parallel, or in different orders. Some may prove unnecessary. Others may divide into several distinct mechanisms.

The scientific value of the sequence lies precisely in its decomposability.

Each transition can in principle fail.

Only after these functional dependencies are understood should a stronger consciousness hypothesis be considered:

Do particular classes of such organizational dynamics constitute necessary, enabling, or sufficient conditions for phenomenal consciousness?

At present, none of these alternatives is established.

11. Future Meta-Evolution Research: A Dependency DAG

The next scientifically useful step is therefore not to introduce another definition of consciousness.

Instead, the problem should be decomposed into an explicit dependency DAG separating constructs that are often compressed into the intuitive category of “self-aware consciousness.”

The initial node set should include:

Phenomenal Consciousness / Minimal Self / Narrative Self / DMN-Related Integration / Self-Model / Metacognition / Causal Access to Adaptive Operators / Reflective Evolution

The purpose of this DAG is not to impose a linear hierarchy among these constructs. Its purpose is to determine which dependencies are empirically supported, which are merely correlational, which require additional bridge assumptions, and which would constitute new hypotheses within Meta-Evolution.

For example, the current literature provides substantial empirical support for associations between DMN organization and autobiographical memory, self-reference, internally generated cognition, prospection, and narrative integration (Buckner & DiNicola, 2019; Menon, 2023).

There is also evidence that perturbations affecting ordinary self-experience, including meditation and psychedelic ego dissolution, are associated with substantial changes in DMN activity and connectivity while conscious experience remains present (Brewer et al., 2011; Siegel et al., 2024).

Conversely, severe disorders of consciousness are associated with disruptions of DMN connectivity, indicating that the network cannot simply be treated as an optional mechanism of autobiographical narration with no relevance to the broader organization supporting conscious cognition (Fernández-Espejo et al., 2012).

Taken together, however, these findings do not establish a simple causal sequence such as:

DMN → Self → Consciousness.

In particular, evidence that DMN organization contributes to self-related processing does not imply that the DMN generates the self, and evidence linking DMN integrity to conscious state does not establish that the DMN generates phenomenal consciousness.

The appropriate DAG should therefore distinguish at least three different questions:

DMN-Related Integration → Narrative Self

Narrative Self ↔ Minimal Self

Minimal Self / Narrative Self → Phenomenal Consciousness

The first relation is empirically motivated but should not be interpreted as a one-to-one causal identity. The second remains theoretically and empirically complex, since minimal and narrative selfhood may interact without being reducible to one another. The third must remain an unresolved bridge hypothesis rather than an assumed dependency.

A similar distinction applies to the Meta-Evolution branch.

The relation

ρ(t, O(t)) → ΔO(t), together with O(t + 1) = Θ(O(t), ΔO(t)),

describes reflective causal access within the Meta-Evolution framework. Here, ρ(t, O(t)) denotes a functionally accessible representation of the current adaptive operator, ΔO(t) denotes a modification generated through processes that make use of this representation, and Θ specifies how the current operator and its modification determine the adaptive operator governing subsequent dynamics.

This relation defines a pathway from representation of an adaptive operator to modification of that operator. It does not establish any necessary relation between Reflective Evolution and phenomenal consciousness.

Accordingly, the DAG must not silently introduce the inference:

Reflective Evolution → Phenomenal Consciousness.

Nor should the reverse inference be assumed:

Phenomenal Consciousness → Reflective Evolution.

A system could, in principle, possess phenomenal consciousness without functionally representing or modifying its own adaptive operators. Conversely, a sufficiently advanced artificial adaptive system could potentially satisfy the formal requirements of Reflective Evolution without providing independent evidence of phenomenal experience.

The Meta-Evolution branch should therefore initially be represented as:

Self-Model → Metacognitive Access → Causal Access to Adaptive Operators → Reflective Evolution,

with each arrow treated as a candidate dependency requiring separate operational justification.

Even this sequence should not be assumed to be universally necessary. A system may possess metacognitive access to selected internal processes without maintaining a rich narrative self-model, and causal access to adaptive operators may in principle be implemented through mechanisms that differ substantially from human metacognition.

The central methodological principle is therefore:

no edge in the DAG should be accepted merely because two constructs are conceptually adjacent.

Each edge must receive an explicit epistemic status.

A useful initial classification would distinguish:

Empirically Supported Relation

A dependency or association supported by convergent experimental evidence.

Correlational Association

A reproducible relationship for which causal direction has not been established.

Bridge Hypothesis

A proposed relation connecting constructs defined at different explanatory or observational levels and requiring independent validation.

Meta-Evolution Extension Hypothesis

A new organizational proposition introduced within the Meta-Evolution research program and therefore subject to its formal, computational, and empirical validation requirements.

The resulting DAG would therefore function not as a diagram of an established theory, but as an epistemic map of the research problem.

Figure 5. Candidate epistemic dependency DAG. Solid, dashed, and dotted edges distinguish better-supported relations from candidate dependencies and unresolved bridge hypotheses.

Its central task would be to determine which paths, if any, connect:

Phenomenal Consciousness

Minimal Self

Narrative Self

DMN-Related Integration

Self-Model

Metacognition

Causal Access to Adaptive Operators

Reflective Evolution.

Only after these relations have been individually tested should stronger composite hypotheses be considered.

This prevents the emergence of an unsupported chain such as:

DMN → Self → Metacognition → Reflective Evolution → Consciousness.

Instead, each transition becomes an independent scientific question.

In this sense, the dependency DAG would provide the methodological interface between contemporary consciousness research and a possible future extension of Meta-Evolution.

A likely early version would treat the link between DMN-related integration and narrative self as empirically well motivated but not reducible to a single network mechanism.

The connection between predictive interoceptive regulation and embodied minimal self would be treated as a theoretically developed and empirically motivated relation rather than an established identity.

The connection between metacognition and explicit self-awareness would require construct-specific operational definitions.

The connection between a self-model and causal access to adaptive operators would constitute a candidate organizational bridge.

The transition from causal operator access to Reflective Evolution would follow the existing ME definition.

Most importantly, all proposed edges from these functional constructs to phenomenal consciousness would initially remain unresolved bridge hypotheses.

This architecture directly follows the methodological discipline already present in Meta-Evolution: conceptual distinction first, operational reconstruction second, validation of bridges third, and formal extension only after those steps justify it.

The corresponding computational program could then construct artificial agents with progressively richer architectures while controlling individual components: persistent memory, self-state estimation, prospective simulation, metacognitive access, internally generated offline processing, and adaptive-operator modification.

Ablation would be essential.

If removing a self-model does not affect long-horizon autonomous control, then its proposed organizational function has been overstated.

If metacognition produces no independent predictive advantage after controlling for memory and planning, its role must be revised.

If access to operator representations does not causally improve or alter operator adaptation, the criterion for Reflective Evolution has not been met.

Such results would be scientifically valuable even if no claim concerning consciousness could be made.

The central experimental question would therefore become:

Does increasing adaptive autonomy generate reproducible transitions toward persistent self-modelling, temporally extended intrinsic integration, metacognitive access, and ultimately reflective control of adaptive operators?

Only after that dependency structure is established should the stronger question be reopened:

Are any of these transitions systematically associated with the conditions under which phenomenal consciousness arises?

12. Conclusions

The question of artificial consciousness may be difficult partly because several distinct phenomena have historically been compressed into a single concept.

Conscious experience is not identical to autobiographical identity.

Autobiographical identity is not identical to minimal selfhood.

Self-reference is not identical to metacognition.

Metacognition is not identical to causal access to adaptive mechanisms.

Reflective Evolution is not identical to consciousness.

The Default Mode Network provides an especially valuable biological example of this separation. Research increasingly associates DMN-related systems with internally generated representation, autobiographical memory, semantic integration, prospection, social cognition, and construction of a coherent internal narrative. It is therefore highly relevant to the organization of human selfhood. Yet current evidence does not justify treating the DMN as the neural location or generator of consciousness itself.

Meditation, psychedelic ego dissolution, and research on minimal phenomenal experience further indicate that ordinary narrative selfhood can be strongly altered without eliminating conscious experience. At the same time, disorders-of-consciousness research shows that DMN integrity participates in the broader network conditions associated with normal conscious cognition.

The resulting picture is neither “the ego creates consciousness” nor “the ego is irrelevant.”

Instead, conscious biological systems appear to contain several nested and interacting organizational layers.

For artificial systems, the most informative analogue of the DMN would therefore not be a particular artificial neural layer. It would be a persistent intrinsic integration regime in which memories, world representations, self-state estimates, possible futures, goals, other agents, and long-term constraints are continuously combined into a model capable of influencing subsequent behaviour.

Such an artificial architecture could acquire something functionally analogous to a narrative self.

It could perhaps acquire a minimal regulatory self-model.

It could become metacognitive.

It could eventually obtain causal access to its own adaptive operators and therefore satisfy the Meta-Evolution criterion for Reflective Evolution.

None of these transitions, individually or jointly, presently proves phenomenal consciousness.

This negative result is scientifically productive.

It relocates the central problem.

The crucial boundary may not lie between carbon and silicon, or even between biological and artificial intelligence. It may lie among different forms of causal organization: systems that merely process external information; systems that model the world; systems that model themselves within that world; systems whose own future continuity becomes a persistent variable of prediction and control; and systems capable of modifying the mechanisms that determine how they will adapt in the future.

Whether phenomenal consciousness emerges at one of these transitions, requires a combination of them, depends upon additional biological mechanisms, or belongs to an entirely different explanatory level remains unknown.

The appropriate next step is therefore not to declare artificial systems conscious or unconscious from behavioural resemblance.

It is to construct the dependency structure.

A rigorous DAG connecting phenomenal consciousness, minimal self, narrative self, DMN-related integration, self-modeling, metacognition, causal operator access, and Reflective Evolution would transform a broad ontological debate into a sequence of separable hypotheses.

Some edges are already empirically constrained.

Some are correlations.

Some require bridge models.

Some would constitute genuinely new extensions of Meta-Evolution.

This distinction creates a scientific route forward while preserving the central epistemic limitation:

we may be able to explain increasingly sophisticated forms of self-organization before we can explain why any organization is accompanied by experience at all.

13. References

Brewer, J. A., Worhunsky, P. D., Gray, J. R., Tang, Y.-Y., Weber, J., & Kober, H. (2011). Meditation experience is associated with differences in default mode network activity and connectivity. Proceedings of the National Academy of Sciences, 108(50), 20254–20259. doi: 10.1073/pnas.1112029108.

Buckner, R. L., & DiNicola, L. M. (2019). The brain’s default network: Updated anatomy, physiology and evolving insights. Nature Reviews Neuroscience, 20, 593–608. doi: 10.1038/s41583-019-0212-7.

Butlin, P., Long, R., Bayne, T., Bengio, Y., Birch, J., Chalmers, D., Constant, A., Deane, G., Elmoznino, E., Fleming, S. M., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2026). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences, 30(6), 488–501. doi: 10.1016/j.tics.2025.10.011.

Cogitate Consortium, Ferrante, O., Gorska-Klimowska, U., et al. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature, 642, 133–142. doi: 10.1038/s41586-025-08888-1.

Fernández-Espejo, D., Soddu, A., Cruse, D., Palacios, E. M., Junque, C., Vanhaudenhuyse, A., Rivas, E., Newcombe, V., Menon, D. K., Pickard, J. D., Laureys, S., & Owen, A. M. (2012). A role for the default mode network in the bases of disorders of consciousness. Annals of Neurology, 72(3), 335–343. doi: 10.1002/ana.23635.

Gallagher, S. (2000). Philosophical conceptions of the self: Implications for cognitive science. Trends in Cognitive Sciences, 4(1), 14–21. doi: 10.1016/S1364-6613(99)01417-5.

Gusnard, D. A., Akbudak, E., Shulman, G. L., & Raichle, M. E. (2001). Medial prefrontal cortex and self-referential mental activity: Relation to a default mode of brain function. Proceedings of the National Academy of Sciences, 98(7), 4259–4264. doi: 10.1073/pnas.071043098.

Hoffman, J. (2026). Meta-Evolution: An Independent Research Program on Adaptive Information and Higher-Order Evolution. Version 1.97. Zenodo. doi: 10.5281/zenodo.21722894.

Menon, V. (2023). 20 years of the default mode network: A review and synthesis. Neuron, 111(16), 2469–2487. doi: 10.1016/j.neuron.2023.04.023.

Metzinger, T. (2024). The Elephant and the Blind: The Experience of Pure Consciousness: Philosophy, Science, and 500+ Experiential Reports. MIT Press. doi: 10.7551/mitpress/15196.001.0001.

Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, Article 2, 1–22. doi: 10.1145/3586183.3606763.

Qin, P., & Northoff, G. (2011). How is our self related to midline regions and the default-mode network? NeuroImage, 57(3), 1221–1233. doi: 10.1016/j.neuroimage.2011.05.028.

Raichle, M. E., MacLeod, A. M., Snyder, A. Z., Powers, W. J., Gusnard, D. A., & Shulman, G. L. (2001). A default mode of brain function. Proceedings of the National Academy of Sciences, 98(2), 676–682. doi: 10.1073/pnas.98.2.676.

Seth, A. K., & Bayne, T. (2022). Theories of consciousness. Nature Reviews Neuroscience, 23, 439–452. doi: 10.1038/s41583-022-00587-4.

Seth, A. K., & Hohwy, J. (2021). Predictive processing as an empirical theory for consciousness science. Cognitive Neuroscience, 12(2), 89–90. doi: 10.1080/17588928.2020.1838467.

Seth, A. K., & Tsakiris, M. (2018). Being a Beast Machine: The somatic basis of selfhood. Trends in Cognitive Sciences, 22(11), 969–981. doi: 10.1016/j.tics.2018.08.008.

Siegel, J. S., Subramanian, S., Perry, D., Kay, B. P., Gordon, E. M., Laumann, T. O., et al. (2024). Psilocybin desynchronizes the human brain. Nature, 632, 131–138. doi: 10.1038/s41586-024-07624-5.

Yeshurun, Y., Nguyen, M., & Hasson, U. (2021). The default mode network: Where the idiosyncratic self meets the shared social world. Nature Reviews Neuroscience, 22, 181–192. doi: 10.1038/s41583-020-00420-w.

Research context

This essay is part of the broader Meta-Evolution research program presented in the full reference monograph:

Jacek Hoffman, Meta-Evolution: An Independent Research Program on Adaptive Information and Higher-Order Evolution v.1.97

The monograph presents the formal architecture of the program, operational definitions, the Meta-Evolution Observation Protocol (MEOP), computational and experimental research proposals, and nine candidate falsification programs across biological, medical, artificial, environmental, and institutional systems.

Access the complete research program:

· Open-access reference edition: PDF and DOI

· Amazon Kindle edition

Research context

This essay is part of the broader Meta-Evolution research program presented in the full reference monograph:

Jacek Hoffman, Meta-Evolution: An Independent Research Program on Adaptive Information and Higher-Order Evolution.

The monograph presents the formal architecture of the program, operational definitions, the Meta-Evolution Observation Protocol (MEOP), computational and experimental research proposals, and nine candidate falsification programs across biological, medical, artificial, environmental, and institutional systems.

Access the complete research program:

• Open-access reference edition: PDF and DOI

• Amazon Kindle edition

Preface

This essay takes Terry Sejnowski’s article Part 20: Schizophrenia: Searching for the Prime Suspect (August 2, 2026), together with the related Brain & Behavior Research Foundation discussion of convergent brain-network changes in schizophrenia, as a starting point for further analysis. Rather than treating the PV-interneuron hypothesis as a final explanation, the essay develops a broader interpretation of schizophrenia as a disorder of predictive processing, stress-dependent learning, salience attribution, and the stabilization of threat-related models of reality.

In my view, schizophrenia may be driven by a disruption of the brain’s predictive system.

The proposal that schizophrenia involves impaired regulation and synchronization of cortical networks is persuasive. It moves the discussion beyond the traditional search for an excess or deficiency of a single neurotransmitter. However, it does not by itself establish a uniquely privileged role for parvalbumin-positive, or PV, interneurons. Their dysfunction may be important, but it is more plausibly treated as one component of a broader disturbance affecting the brain’s adaptive and regulatory architecture.

Post-mortem and physiological evidence connects PV-cell alterations with weakened GABAergic inhibition, disturbed gamma oscillations, and cognitive impairment [1,2]. These findings support the relevance of PV interneurons without demonstrating that they constitute the unique origin of schizophrenia.

The similarity between schizophrenia and transient ketamine-induced psychosis should therefore be interpreted within this broader context. Ketamine may reveal how disruption of NMDA-dependent regulation can destabilize cortical inhibition, synchronization, and information processing. It does not necessarily identify a single cause of schizophrenia [2].

The analogy with Parkinson’s disease is likewise heuristically useful but limited. Parkinson’s disease is associated with a relatively well-defined degeneration of dopaminergic neurons, whereas schizophrenia is far more heterogeneous. It includes different clinical courses, symptom profiles, cognitive impairments, and probable biological mechanisms.

A broader interpretation begins with hierarchical predictive inference.

The brain continuously generates predictions about the causes of its sensory input, compares them with incoming information, and updates its internal model. This process depends not only on prediction errors, but also on the precision assigned to them. The system must estimate which signals are reliable, which are noise, and which require a revision of its model of reality.

Experimental findings have connected psychosis and delusion formation with disturbed prediction-error processing and abnormal precision weighting [3,4]. The aberrant-salience hypothesis similarly proposes that dopaminergic dysregulation can cause neutral or irrelevant events to acquire excessive motivational and explanatory significance [5].

Figure 1. Conceptual scheme of hierarchical predictive inference. Precision weighting determines how strongly prediction errors modify the internal model. The figure is schematic and does not represent measured data.

This is where the interaction between dopamine and cortisol acquires a deeper evolutionary meaning. They can be understood as components of a coupled reinforcement and correction system, broadly analogous to reinforcement-learning architectures.

Dopamine participates in rapid signalling related to prediction error, expected value, motivation, and salience. Cortisol regulates the wider adaptive state within which these signals are interpreted.

From an evolutionary perspective, cortisol is part of an allostatic survival system. In an aggressive, unstable, or unpredictable environment, the organism must rapidly mobilize energy, redirect attention, modify memory formation, and adjust its behavioural strategy. When this response is temporary and efficiently terminated, it supports adaptation. When it is repeatedly activated or poorly regulated, the same mechanism can generate allostatic load and contribute to pathology [6,7].

The evolutionary logic of this system is based on an asymmetry of errors. Mistaking a harmless sound for a predator wastes energy. Mistaking a predator for a harmless sound may end the organism’s life. Under conditions of danger, the adaptive system should therefore become more sensitive to weak indications of threat and more willing to produce false alarms.

In computational terms, cortisol may be interpreted as modifying the cost landscape of learning:

δ_t = r_t − λ_t c_t + γV(s_{t+1}) − V(s_t)

Here, t denotes the current time step or decision episode. The term δt represents the prediction error at time t, that is, the discrepancy between expected and obtained value. The quantity rt denotes the immediate reward or positive outcome obtained at time t. The term ct represents a possible cost, loss, or threat associated with the current situation. The coefficient λt determines how strongly this cost is weighted at time t; in the present interpretation, it captures the degree to which potential threat or loss influences learning and decision-making. The factor γ is the temporal discount parameter, which specifies how strongly the system values future consequences relative to immediate ones, with 0≤γ≤1. The function V(st) denotes the estimated value of the current state st, while V(st+1) denotes the estimated value of the next state st+1.

Under this interpretation, an increase in λt would make the system more sensitive to potential losses, more cautious when facing uncertain alternatives, and more likely to treat ambiguous signals as possible warnings. Cortisol does not literally compute a scalar cost function. Rather, it may biologically alter the conditions under which learning occurs by changing sensitivity to risk, attention to threat, behavioural flexibility, and the balance between model-based and habitual control [8].

Dopamine and cortisol can therefore be described as functional “plus” and “minus” elements only at a high level of abstraction. Dopamine provides rapid information about changes in expected value and significance. Cortisol changes the broader learning regime by increasing the importance of possible danger, loss, and uncertainty.

The crucial point is that cortisol does not need to remain permanently elevated in order to contribute to a persistent pathological state.

Its decisive role may occur during the formation, consolidation, or later reconsolidation of a predictive pattern. When an experience is encoded under conditions of intense threat, cortisol and associated arousal systems can increase the adaptive weight assigned to that experience. A temporary hormonal state may therefore produce a lasting change in the brain’s predictive model [9].

acute threat ⟶ cortisol and arousal ⟶ increased weight of the experience ⟶ consolidation of a threat model

Once this model has been consolidated, cortisol may return to baseline while the altered predictive structure remains. The organism has learned a rule:

the environment is dangerous, unpredictable, and personally significant

Ambiguous signals are subsequently interpreted through this previously acquired threat model. Attention is directed toward confirming evidence, while contradictory information may be ignored, discounted, or reinterpreted.

Each new episode of fear can reactivate the established pattern. During reconsolidation, the model may be strengthened and written back into memory. An initially adaptive response to a real danger can therefore become a stable pathological attractor.

Figure 2. Conceptual threat-learning loop. A transient stress response may influence consolidation; the resulting threat prior can persist and be strengthened through salience assignment, avoidance, reactivation, and reconsolidation. The figure is a theoretical synthesis, not a direct empirical measurement.

The process may develop as follows:

neural noise ⟶ incorrectly weighted prediction errors ⟶ excessive salience or threat attribution ⟶ significance assigned to random events

If neutral stimuli repeatedly generate the signal “this is important,” the brain must infer a cause. In a social environment, the most readily available cause is often another person’s intention.

The inference can gradually move from:

something important is happening ⟶ someone is causing it ⟶ someone is acting against me

A persecutory delusion may then emerge as an attempt to organize chaotic and contradictory prediction errors. It is not necessarily an arbitrary belief. Initially, it may provide a coherent explanation for an increasingly unstable and threatening world.

Once established, however, the delusion can become resistant to correction. New events are interpreted through the explanatory model it has created. Avoidance reduces access to corrective information, while selective attention supplies apparent confirmation.

The resulting loop is self-reinforcing:

expectation of threat ⟶ selective attention ⟶ confirming signals ⟶ avoidance ⟶ reduced corrective experience ⟶ reinforcement

In this sense, schizophrenia, especially in its persecutory or paranoid dimension, may involve a relatively persistent state of inferred threat. This does not require permanently elevated cortisol. The hormonal signal may have played its most important role earlier, during the learning and consolidation of the pathological prediction.

Dopaminergic salience mechanisms may subsequently mark events as important, while the consolidated threat model determines how that importance is interpreted. A sound, glance, coincidence, or fragment of conversation becomes meaningful because the system has already learned that weak signals may indicate hidden danger.

Hallucinations can be understood through a related imbalance. Internally generated predictions may receive too much weight relative to sensory evidence. Conditioning experiments suggest that susceptibility to hallucinations can be associated with excessive reliance on perceptual priors [10].

A disturbance of efference copy or corollary discharge may further weaken the distinction between self-generated and externally generated signals. If the sensory consequences of inner speech or voluntary action are not correctly predicted, thoughts and intentions may be experienced as originating from an external agent [11].

Within this framework, the PV-cell hypothesis does not compete with the predictive-processing hypothesis. It describes one possible biological substrate of predictive dysregulation.

PV interneurons regulate gamma synchronization and the millisecond-scale temporal windows through which cortical signals reinforce, coordinate, or suppress one another. Their dysfunction may impair temporal precision, signal-to-noise ratio, and the encoding of informational reliability [1,2].

The weakening of PV-mediated inhibition does not need to produce a simple increase in overall brain activity. Its more important consequence may be a loss of precision in determining which signals are reliable, which are internally generated, and which deserve behavioural significance.

The stress system, dopamine, NMDA receptors, GABAergic inhibition, gamma oscillations, and cortical connectivity can therefore be understood as interacting levels of the same process.

Evidence also suggests that the normal relationship between cortisol responses and prefrontal dopamine release during stress may be altered in schizophrenia [12]. This does not prove that cortisol initiates the disorder, but it supports the possibility that the systems responsible for stress adaptation and salience assignment become functionally uncoupled or incorrectly coordinated.

Figure 3. Multilevel interpretation of predictive dysregulation. Molecular, cellular, circuit, dynamic, computational, and clinical descriptions are treated as interacting levels rather than competing explanations.

Schizophrenia may therefore be a heterogeneous disorder with a partially shared computational core: dysregulation of hierarchical predictive processing, including the miscoding of precision, salience, threat, and authorship.

In this model, psychosis develops when a system designed to learn rapidly from danger begins to detect danger, intention, and meaning where the available evidence is too weak to justify them.

A temporary state of biological alarm may help establish the pathological prediction. Once encoded, the prediction can persist through neural plasticity, selective attention, avoidance, dopaminergic salience, and repeated reconsolidation.

The brain no longer merely reacts to a dangerous world. It has learned to predict one.

References

1. Curley, A. A., & Lewis, D. A. (2012). Cortical basket cell dysfunction in schizophrenia. The Journal of Physiology, 590(4), 715–724. https://doi.org/10.1113/jphysiol.2011.224659

2. Jadi, M. P., Behrens, M. M., & Sejnowski, T. J. (2016). Abnormal gamma oscillations in N-methyl-D-aspartate receptor hypofunction models of schizophrenia. Biological Psychiatry, 79(9), 716–726. https://doi.org/10.1016/j.biopsych.2015.07.005

3. Corlett, P. R., Murray, G. K., Honey, G. D., et al. (2007). Disrupted prediction-error signal in psychosis: Evidence for an associative account of delusions. Brain, 130(9), 2387–2400. https://doi.org/10.1093/brain/awm173

4. Haarsma, J., Fletcher, P. C., Griffin, J. D., et al. (2021). Precision weighting of cortical unsigned prediction error signals benefits learning, is mediated by dopamine, and is impaired in psychosis. Molecular Psychiatry, 26, 5320–5333. https://doi.org/10.1038/s41380-020-0803-8

5. Kapur, S. (2003). Psychosis as a state of aberrant salience: A framework linking biology, phenomenology, and pharmacology in schizophrenia. American Journal of Psychiatry, 160(1), 13–23. https://doi.org/10.1176/appi.ajp.160.1.13

6. McEwen, B. S. (1998). Stress, adaptation, and disease: Allostasis and allostatic load. Annals of the New York Academy of Sciences, 840, 33–44. https://doi.org/10.1111/j.1749-6632.1998.tb09546.x

7. McEwen, B. S. (2007). Physiology and neurobiology of stress and adaptation: Central role of the brain. Physiological Reviews, 87(3), 873–904. https://doi.org/10.1152/physrev.00041.2006

8. Otto, A. R., Raio, C. M., Chiang, A., Phelps, E. A., & Daw, N. D. (2013). Working-memory capacity protects model-based learning from stress. Proceedings of the National Academy of Sciences, 110(52), 20941–20946. https://doi.org/10.1073/pnas.1312011110

9. Roozendaal, B., McEwen, B. S., & Chattarji, S. (2009). Stress, memory and the amygdala. Nature Reviews Neuroscience, 10(6), 423–433. https://doi.org/10.1038/nrn2651

10. Powers, A. R., Mathys, C., & Corlett, P. R. (2017). Pavlovian conditioning-induced hallucinations result from overweighting of perceptual priors. Science, 357(6351), 596–600. https://doi.org/10.1126/science.aan3458

11. Ford, J. M., & Mathalon, D. H. (2004). Electrophysiological evidence of corollary discharge dysfunction in schizophrenia during talking and thinking. Journal of Psychiatric Research, 38(1), 37–46. https://doi.org/10.1016/S0022-3956(03)00095-5

12. Schifani, C., Tseng, H. H., Kenk, M., et al. (2018). Cortical stress regulation is disrupted in schizophrenia but not in clinical high risk for psychosis. Brain, 141(7), 2213–2224. https://doi.org/10.1093/brain/awy133

Research context

This essay is part of the broader Meta-Evolution research program presented in the full reference monograph:

Jacek Hoffman, Meta-Evolution: An Independent Research Program on Adaptive Information and Higher-Order Evolution.

The monograph presents the formal architecture of the program, operational definitions, the Meta-Evolution Observation Protocol (MEOP), computational and experimental research proposals, and nine candidate falsification programs across biological, medical, artificial, environmental, and institutional systems.

Access the complete research program:

• Open-access reference edition: PDF and DOI

• Amazon Kindle edition

1. Starting Point

The human brain is one of the most complex structures known in nature.

It contains approximately 86 billion neurons and hundreds of trillions of synapses, consumes about 20 watts of power, represents only about 2% of body mass, yet accounts for roughly 20% of the body’s resting energy consumption.

This raises a fundamental question:

Can intelligence continue to increase through the further evolution of the individual brain?

If so, by what mechanism?

2. The First Strategy — More Neurons

The most intuitive solution appears straightforward.

More neurons.

A larger brain.

Greater intelligence.

However, this approach immediately encounters several fundamental constraints.

More neurons require:

• higher energy consumption,

• greater oxygen and glucose supply,

• increased heat production,

• a larger cranial volume,

• longer axons,

• longer communication delays.

More importantly, one cannot simply enlarge the brain.

Virtually the entire organism would have to change:

• the cardiovascular system,

• the respiratory system,

• metabolic capacity,

• thermoregulation,

• skull morphology,

• the birth canal,

• embryonic development,

• the duration of childhood and neural maturation.

Increasing neuron number therefore ceases to be a local modification of the brain.

It becomes a redesign of the entire organism.

3. The Second Strategy — More Connections

An alternative would be to keep approximately the same number of neurons while increasing the number of synaptic connections.

This strategy also faces fundamental limitations.

More synapses mean:

• higher maintenance costs,

• greater synchronization costs,

• increased neural noise,

• more complex communication,

• higher learning and plasticity costs.

Beyond a certain point, the cost of communication begins to exceed the computational benefit provided by additional connections.

A similar phenomenon is well known in computer science.

As the number of processors increases, computation itself eventually becomes relatively inexpensive, while communication and synchronization become the dominant bottlenecks.

4. Biological Limits

Both strategies lead to the same conclusion.

The number of neurons and the number of synapses can increase only up to a certain point.

Beyond that point, costs grow disproportionately:

• energetic,

• metabolic,

• communicational,

• developmental,

• evolutionary.

It is therefore plausible that the individual biological brain is approaching the limits of its evolutionary scalability.

This does not imply that intelligence has reached its maximum.

It only suggests that further increases in intelligence may require a fundamentally different mechanism.

An Evolutionary Observation — There Is No Single Optimal Brain

The human brain is often regarded as the pinnacle of biological intelligence.

However, evolution has produced several remarkably different nervous system architectures.

Cetaceans, including dolphins and killer whales, possess brains larger than those of humans and exhibit highly folded cerebral cortices together with sophisticated social cognition. Yet despite their larger brains, there is no evidence that they surpass humans in general intelligence. Much of their neural complexity appears adapted to life in the aquatic environment, including acoustic processing, echolocation, and large-body sensorimotor control.

Octopuses represent an even more striking example.

Rather than concentrating computation in a single central brain, approximately two-thirds of their neurons are distributed throughout their arms, allowing substantial local information processing.

These examples suggest an important conclusion.

Evolution has not converged on a single optimal brain architecture. Instead, it has repeatedly explored different ways of organizing information processing while achieving high adaptive intelligence.

Consequently, neuron number alone cannot explain cognitive capability.

The organization of information processing appears to be equally important.

A Different Perspective on Brain Evolution

If multiple nervous system architectures can produce sophisticated intelligence, the central evolutionary question changes.

Instead of asking:

How can evolution build an even larger brain?

one may ask:

What organizational architectures allow information processing to scale beyond the limits of an individual brain?

This shift is fundamental.

It suggests that future evolutionary innovation may occur not by indefinitely increasing the size or complexity of individual brains, but by reorganizing how multiple cognitive systems interact.

From a Meta-Evolutionary perspective, the next adaptive transition may therefore involve a change in the unit of cognition itself.

5. A Lesson from Evolution

The diversity of biological nervous systems illustrates a broader evolutionary principle.

Whenever an existing organizational architecture approaches its practical limits, evolution often responds not by indefinitely optimizing that architecture, but by creating a new organizational level.

Examples include:

• molecules → cells,

• cells → multicellular organisms,

• organisms → eusocial societies.

Each transition increased adaptive capacity not by endlessly improving individual components, but by reorganizing them into a higher-level system.

The question is whether the biological brain represents another such transition point.

6. A Meta-Evolutionary Hypothesis

If this pattern is general, one can formulate the following hypothesis.

If the individual brain approaches the limits of further scaling, evolution is unlikely to continue increasing either neuron number or synaptic density indefinitely. Instead, it becomes more probable that a new level of organization will emerge, integrating multiple brains into a higher-order cognitive system.

In such a framework, the primary cognitive unit is no longer the individual organism.

It becomes a network of interacting cognitive systems.

7. Channels of Integration

Within this framework, the critical factor is no longer brain size.

The crucial variable becomes the communication channel connecting individual cognitive systems.

Viewed from this perspective, the history of civilization can be interpreted as the history of increasingly powerful integration channels.

Language enabled the transfer of knowledge.

Writing enabled durable storage.

Printing multiplied dissemination.

The Internet reduced communication delays to nearly zero.

Artificial intelligence is beginning to function as an integrative layer, translating concepts, compressing knowledge, and reconstructing relationships across domains.

Rather than increasing biological intelligence directly, it increases effective Integrative Cognitive Bandwidth by reducing the cost of knowledge integration.

Future developments may include:

• brain–computer interfaces,

• brain–brain interfaces,

• new communication technologies,

• and perhaps—if physically possible—novel biological communication mechanisms.

The particular implementation, however, is secondary.

8. The Underlying Mechanism

Regardless of implementation, the underlying mechanism remains the same.

We do not primarily improve the individual neuron.

We do not even primarily improve the individual brain.

Instead, we transfer part of the cognitive process to a higher level of organization.

Conceptually,

neurons
      ↓
brain
      ↓
network of brains
      ↓
integrative layer
      ↓
higher-order cognitive system

This represents exactly the same type of organizational transition repeatedly observed throughout evolutionary history.

9. Central Hypothesis

The central hypothesis can therefore be formulated as follows:

If the individual brain approaches the limits of further evolutionary scaling, Meta-Evolution predicts the emergence of increasingly powerful channels of integration between brains. The specific implementation of these channels—language, writing, the Internet, artificial intelligence, brain–computer interfaces, or perhaps, in the distant future, entirely new biological mechanisms—is secondary. The essential mechanism is the transfer of part of the cognitive process to a higher organizational level.

10. A New Scientific Question

This shifts the discussion away from questions such as:

Is telepathy possible?

toward a much broader and scientifically more productive question:

Does the emergence of progressively richer channels of integration between information-processing systems represent a general law of Meta-Evolution?

This formulation does not require assuming any particular communication medium, nor does it depend on speculative or metaphysical hypotheses.

Instead, it focuses on an organizational principle repeatedly observed throughout evolutionary history:

When an existing level of organization approaches its adaptive limits, evolution increases complexity by creating a higher level of organization capable of integrating information more effectively.

11. Testable Predictions

If this hypothesis is correct, it generates several empirical predictions.

One would expect the long-term evolution of civilization to exhibit:

• progressively higher communication bandwidth between cognitive systems,
  

• increasingly efficient semantic compression,
  

• deeper integration of heterogeneous knowledge,
  

• the emergence of stable human–AI cognitive clusters,
  

• the transfer of an increasing fraction of cognitive processing from individuals to distributed networks,
  

• the emergence of organizational architectures whose cognitive capacity exceeds the simple sum of their individual components.
  

If this hypothesis is correct, future evolutionary transitions should be characterized less by increases in the complexity of individual brains and more by increases in the efficiency with which independent cognitive systems exchange, integrate, and coordinate information.

12. Conclusion

The existence of fundamentally different biological solutions to intelligence—including the centralized human brain, the large socially specialized brains of cetaceans, and the distributed nervous system of octopuses—demonstrates that evolution is remarkably flexible in designing cognitive architectures.

This suggests that intelligence is not tied to a single optimal neural design.

Instead, evolution appears to explore multiple organizational architectures capable of producing sophisticated cognition.

If this organizational flexibility extends beyond the boundaries of individual organisms, the emergence of distributed cognitive systems may represent not an anomaly but the next natural transition in the evolution of intelligence.

From this perspective, the evolution of intelligence does not end with the biological brain.

Rather, the biological brain represents another stage in a much longer sequence of organizational transitions whose common function is to increase the capacity for integrating, organizing, and utilizing information.

The central claim of this hypothesis is therefore not about telepathy, artificial intelligence, or any particular technology.

It is about a more general evolutionary principle:

As existing cognitive architectures approach their scaling limits, adaptive evolution shifts from improving individual components toward creating new organizational levels connected by increasingly efficient channels of information integration.

I would go one step further.

If this principle proves to be correct, then the fundamental unit of evolution may not be the individual organism, nor even the individual brain, but the highest level of organization capable of integrating information under the prevailing physical constraints.

That possibility, rather than any particular communication technology, constitutes the core scientific hypothesis of Meta-Evolution.

Further Reading

A broader discussion of evolutionary systems and the Meta-Evolution framework is available here:

Meta-Evolution: An Independent Research Program on Adaptive Information and Higher-Order Evolution

You might as well ask if a vacuum cleaner is an entropy machine. It sounds absurd, but given the current pace at which neuroscience and AI attempt to “solve” intelligence by obsessing over single neurons, it fits perfectly. Since everyone knows what a vacuum cleaner is for, let’s use it as a scalpel to dissect the methodology of modern cognitive science.

From a statistical perspective, “ordering” (the reduction of configurational entropy) is simply the transition of a system from a high-probability macrostate (dust dispersed in a room) to a low-probability one (dust condensed in a bag). The vacuum cleaner acts as a “brute-force” Maxwell’s Demon—it doesn’t select particles by speed; it imposes a brutal geometric restriction: the filter. The information cost of this sorting process is paid by increasing the disorder of the surroundings (waste heat).

And here is the crux: why trying to compare a biological neuron to an artificial one is a similar brand of absurdity.

Because a neuron is not a “unit of intelligence.” It is merely an infrastructure element, an organizational platform—a glorified “dust separator.”

Modern science is committing a classic category error here:

Component Fetishism: We compare “spikes” (biology) with “floats” (AI), forgetting that these are just mediums. It’s like evaluating the quality of literature based on the difference between ink and pixels, while completely ignoring the text.

Ontological Blindness: A biological neuron manages its own entropy within the framework of homeostasis. An ANN neuron is a dead snippet of code that “lives” only because someone (the server farm) has assumed its entire entropic debt.

Attempting to “understand” intelligence by comparing these two types of neurons is a semantic experiment that makes as much sense as debating whether a bagged vacuum or a cyclone system is “better”—while ignoring the fact that both devices serve only to move disorder from point A to point B.

Of course, biological media and silicon differ computationally. It can be measured. But this entire tirade about “Artificial Intelligence” is just an attempt to build increasingly elaborate vacuum bags, hoping that if the bag is large enough, the dust will spontaneously organize itself into consciousness.

The real question is: will we ever stop obsessing over the “vacuuming” and start investigating why the dust exists in the first place?

Is our entire architecture of neural networks just the most complex “dust bag” in history, which we have, in our ignorance, decided to label “intelligence”?

Biological intelligence is commonly associated with wakefulness, during which perception, reasoning, and decision-making occur. Yet an increasing body of evidence suggests that one of the most important phases of cognition takes place while we sleep. Rather than representing a passive interruption of mental activity, sleep is an active biological process during which memories are consolidated, neural representations reorganized, and newly acquired information selectively integrated into existing knowledge.

Within the systems-level framework proposed in this essay, sleep may be interpreted as one of the principal biological mechanisms supporting Integrative Cognitive Bandwidth (ICB).

Integrative Cognitive Bandwidth is conceptually defined as the maximum rate at which a cognitive system can construct, update, and maintain coherent internal models of reality while operating under constraints imposed by uncertainty, energetic resources, communication costs, and biological limitations.

Working memory determines how many pieces of information can be manipulated simultaneously, but its effectiveness depends critically on the organization of long-term knowledge. Sleep appears to increase this organizational efficiency by consolidating important representations, weakening redundant associations, strengthening meaningful connections, and reorganizing distributed neural activity into more coherent cognitive structures. Consequently, sleep does not merely improve memory. It increases the efficiency with which previously acquired knowledge can be incorporated into existing models of reality, effectively reducing the computational cost of future model updates.

This systems-level interpretation is broadly consistent with several well-established findings in contemporary neuroscience. During Slow-Wave Sleep and REM sleep, activity patterns acquired during wakefulness are repeatedly replayed, facilitating the gradual transfer of information from the hippocampus to distributed cortical networks. Rather than functioning as simple storage, this process continuously restructures internal representations of the world.

A complementary mechanism is provided by the Synaptic Homeostasis Hypothesis proposed by Giulio Tononi and Chiara Cirelli. Throughout the day, learning strengthens large numbers of synapses, increasing both energetic cost and neural noise. During sleep, synaptic strengths are globally downscaled while preserving the most informative connections. The result is a lower metabolic cost, a higher signal-to-noise ratio, and improved learning efficiency during subsequent wakefulness. Within the ICB framework, this process can be interpreted as reducing the cost of maintaining coherent internal representations.

Sleep also supports the molecular foundations of cognition. Sleep deprivation reduces long-term potentiation, disrupts long-term depression, and alters the expression of numerous genes involved in synaptic plasticity, including BDNF, CREB, Arc, and c-Fos. These molecular mechanisms determine the brain’s ability to modify its own circuitry and therefore directly influence its capacity to update internal models of reality.

Another important process is the recently discovered glymphatic system. During sleep, this system removes metabolic waste products, including beta-amyloid and tau proteins. Although glymphatic clearance does not directly increase cognitive performance, it lowers the biological cost of maintaining neural networks, thereby supporting the long-term computational capacity of the brain.

Sleep also affects higher cognitive functions. Even a single night of sleep deprivation produces measurable declines in working memory, executive control, cognitive flexibility, creativity, and the ability to integrate information originating from different conceptual domains. Rather than interpreting these impairments as isolated deficits, the ICB framework suggests viewing them as manifestations of a reduced capacity for model integration.

The same perspective naturally extends to the Default Mode Network, whose activity is profoundly reorganized during sleep. Because this network is strongly associated with autobiographical memory, self-representation, planning, and mental simulation, it provides a plausible biological substrate for the dynamic self-model introduced earlier in this essay.

Taken together, these apparently diverse biological mechanisms reveal a common systems-level organization:

Diagram

Sleep
   |
   +-- Glymphatic clearance
   +-- Gene expression
   +-- Synaptic homeostasis
   +-- Neuroplasticity
   +-- Memory consolidation
   +-- Working memory efficiency
   +-- Creativity
   +-- Self-model stabilization
           |
           v
Integrative Cognitive Bandwidth

Importantly, neuroscience does not currently define or measure Integrative Cognitive Bandwidth. The concept proposed here should therefore be understood as a theoretical framework rather than an established biological quantity. Modern neuroscience describes individual mechanisms such as memory consolidation, synaptic homeostasis, neuroplasticity, glymphatic clearance, and large-scale network dynamics. The ICB framework proposes that these mechanisms may all contribute to a common computational objective: maintaining the brain’s capacity to efficiently integrate heterogeneous information into coherent internal models.

This distinction between empirical observation and theoretical interpretation is essential. The biological mechanisms themselves are well established. Their interpretation as components of a unified systems-level quantity remains the central hypothesis of this work.

From this perspective, the adaptive role of sleep extends beyond improving memory or learning performance. Memory consolidation becomes one component of a broader computational objective: periodically restoring the brain’s ability to integrate knowledge by reorganizing its internal models of reality.

If intelligence is defined primarily as the continuous construction, refinement, and maintenance of coherent models of reality, then sleep should not be viewed merely as a physiological necessity. It is one of the fundamental computational maintenance processes that allows biological intelligence to preserve and periodically restore its Integrative Cognitive Bandwidth.

Within the Entropy Islands framework, this maintenance cycle contributes to the stability of local cognitive organizations, or “Entropy Islands,” enabling biological intelligence to scale while minimizing the synchronization costs associated with increasingly complex internal models.

Further reading

A broader discussion of the Entropy Islands framework, Integrative Cognitive Bandwidth, cognitive clusters, and the role of AI as an integrative layer is available here:

Entropy Islands and the Future of Intelligence

Abstract

This work proposes a conceptual framework linking information theory, computational psycholinguistics, predictive neuroscience, and complex systems theory. The central hypothesis is that natural language should be understood not merely as a communication system but as a long-term training environment for the biological predictive system.

The proposed framework does not claim that one language is inherently superior to another, nor does it argue that language determines intelligence. Instead, it suggests that different languages generate different distributions of predictive costs. Continuous exposure to these distributions throughout development may gradually shape the internal organization of neural representations, producing measurable specialization in selected cognitive operations.

The proposed model is modular, mathematically motivated, and experimentally testable at every stage.

1. Introduction

The discussion of language and cognition has traditionally focused on whether language influences thought. This question has produced decades of debate because it is formulated at a level that is difficult to verify experimentally.

A more productive approach begins elsewhere.

Instead of starting with linguistics or psychology, the present framework begins with information theory and predictive computation.

The central assumption is straightforward.

The brain is viewed as a predictive biological system whose primary function is to minimize prediction error by continuously updating an internal model of the external world.

Within this framework, language is not simply a communication medium. It is one of the largest and most persistent streams of structured data experienced during development. Consequently, language becomes a long-term training operator acting on the predictive machinery of the brain.

2. Language as a Predictive Process

The first stage is not hypothetical.

Natural language is a statistical process.

Modern computational linguistics already describes language using measurable quantities including

• entropy,

• surprisal,

• information density,

• dependency length,

• redundancy,

• locality of syntactic dependencies.

These concepts have been extensively studied for more than two decades.

Experimental evidence consistently demonstrates that language processing cost depends on

• predictability of incoming words,

• distribution of information,

• syntactic dependency length,

• structural integration cost.

Therefore, the first component of the proposed framework is already well established.

No new theory is required at this stage.

3. Predictive Cost Profile

The second step changes the unit of analysis.

Instead of treating language itself as the primary object, each language is viewed as a generator of a characteristic predictive cost profile.

Define

L — a natural language regarded as a statistical process.

Define

Cₚ(L) — the predictive cost profile generated by language L.

The predictive cost profile includes distributions of

• prediction errors,

• dependency lengths,

• information per token,

• integration costs,

• maintenance costs of competing hypotheses,

• uncertainty during incremental parsing.

The crucial observation is that the relevant variable is not language itself.

The important quantity is the predictive cost profile produced by that language.

This interpretation naturally follows from surprisal theory, information density theory, and dependency locality theory.

4. Predictive Problem Space

Once predictive cost profiles are introduced, an important abstraction becomes possible.

Suppose two different languages generate similar predictive cost profiles.

From the perspective of a predictive learning system, both languages represent similar computational problems.

The object of investigation therefore changes.

Instead of studying languages individually, we study a predictive problem space.

Define

P — the space of predictive problems.

Each point inside P represents a class of predictive optimization problems characterized by similar statistical structure.

Individual languages become realizations of particular regions within this space.

This viewpoint substantially simplifies the theoretical framework.

Instead of asking

“How does Polish differ from English?”

we ask

“What class of predictive optimization problem does a language generate?”

The language itself becomes one observable realization of a more general computational object.

5. Adaptation of the Biological Predictive System

Only after defining predictive problem space does neuroscience enter the discussion.

According to predictive processing theories, the brain continuously adjusts its internal representations to minimize expected prediction cost.

Define

E[Cₚ] — expected predictive cost averaged over long-term experience.

The biological learning process attempts to minimize

E[Cₚ]

through continuous adaptation.

Long-term exposure to a particular predictive cost profile therefore produces gradual optimization of internal neural representations.

Several independent areas of neuroscience already support different components of this statement.

Research demonstrates that

• the brain continuously performs prediction,

• language exhibits stable statistical structure,

• surprisal influences neural processing cost,

• dependency length affects working-memory load,

• neural representations remain plastic throughout development.

However, an important theoretical gap still exists.

Current neuroscience does not yet provide a comprehensive theory explaining how long-term exposure to specific predictive cost distributions reorganizes the geometry of neural representations.

This gap represents the central hypothesis of the proposed framework.

6. Neural Representation Geometry

Define

R — the geometry of internal neural representations.

Here, geometry does not refer to physical anatomy.

Instead, it describes the organization of distributed neural representations that encode concepts, relations, predictions, and hierarchical structures.

The central hypothesis can now be stated.

Long-term optimization of predictive cost modifies R.

Changes in R subsequently alter the computational efficiency of selected cognitive operations.

This hypothesis links statistical properties of language directly to representational organization without requiring any assumption about general intelligence.

7. Cognitive Profile

Define

C — the cognitive cost profile of an individual.

This profile represents the computational cost required to perform different classes of cognitive operations.

Examples include

• maintaining long-range dependencies,

• hierarchical reasoning,

• relational memory,

• representation compression,

• model switching,

• structured pattern recognition.

The proposed relationship is

P → R → C

where

• P denotes predictive problem space,

• R denotes neural representation geometry,

• C denotes cognitive cost profile.

The framework does not predict higher intelligence.

Instead, it predicts lower computational cost for specific classes of cognitive operations following prolonged exposure to corresponding predictive problems.

Cognitive specialization is therefore interpreted as computational optimization rather than increased intelligence.

8. Emergent Population-Level Effects

The final stage concerns collective behavior.

No direct causal relationship between language and civilization is proposed.

Instead, define

S — emergent social properties.

These include

• collective problem-solving strategies,

• institutional preferences,

• organizational structures,

• coordination mechanisms.

The hypothesis is probabilistic.

If populations develop statistically different cognitive cost profiles, these differences may influence the probability of adopting particular collective strategies.

Social institutions remain emergent properties of highly complex systems.

Their development depends on geography, ecology, technology, economics, history, education, demography, and numerous additional variables.

Language-derived cognitive specialization, if experimentally confirmed, would represent one contributing factor rather than a dominant cause.

9. Integrated Framework

The complete conceptual model can be summarized as

L → Cₚ(L) → P → R → C → Individual Behavior → S

where

L
: Natural language regarded as a statistical process.

Cₚ(L)
: Predictive cost profile generated by language L.

P
: Predictive problem space defined by families of predictive optimization problems.

R
: Geometry of internal neural representations.

C
: Cognitive cost profile describing computational efficiency across cognitive operations.

S
: Emergent population-level properties resulting from interactions among individuals and environmental constraints.

Each transition corresponds to an independent scientific discipline.

• L → Cₚ(L) belongs primarily to computational psycholinguistics, information theory, and statistical language modeling.

• Cₚ(L) → P belongs to theoretical computational modeling and information geometry.

• P → R belongs to predictive neuroscience and theories of neural plasticity.

• R → C belongs to cognitive neuroscience, experimental psychology, and neuropsychology.

• C → S belongs to complexity science, behavioral economics, quantitative sociology, and evolutionary anthropology.

This modular organization makes every transition independently testable.

Failure of one component would not invalidate the entire framework but would instead identify the precise stage requiring theoretical revision.

10. Conclusions

The proposed framework introduces a different perspective on the relationship between language and cognition.

Rather than asking whether language influences thought, it asks how the statistical structure of language shapes the optimization landscape encountered by a biological predictive system.

Within this framework, language functions as a long-term training operator acting on prediction.

Prediction gradually shapes neural representation geometry.

Representation geometry determines the computational cost of cognitive operations.

Cognitive cost profiles influence individual strategies for solving complex problems.

Finally, interactions among many individuals generate emergent social structures.

The principal contribution of this framework is the shift in the fundamental unit of analysis.

Language is no longer treated primarily as a cultural artifact.

Instead, it is viewed as a generator of predictive cost distributions acting upon an adaptive biological learning system.

This formulation naturally integrates information theory, computational linguistics, predictive neuroscience, learning theory, and complex systems science into a single hierarchical framework that is mathematically motivated, empirically testable, and extensible to future experimental investigation.

Appendix A. Hypothesized Predictive Profiles of Major World Languages

Table A1. Structural language properties and hypothesized predictive specialization.

Methodological note.
The structural language properties listed in this table are well established in descriptive and computational linguistics. The proposed predictive profiles and cognitive adaptations are theoretical hypotheses derived from the framework developed in this paper. They should be regarded as experimentally testable predictions rather than established empirical findings.

Interpretation

Within the proposed framework, individual languages are not considered fundamentally different cognitive systems. Rather, each language represents a particular realization of a predictive optimization problem.

Different structural properties generate different distributions of predictive costs. Long-term exposure to these distributions may gradually bias the optimization of the biological predictive system toward particular classes of computational problems.

The proposed specializations should therefore be interpreted as differences in computational efficiency, not differences in general intelligence.

For example:

• Languages with rich inflection and flexible word order are hypothesized to promote efficient tracking of long-range relational dependencies.

• Languages with rigid word order may favor highly efficient local sequential prediction.

• Agglutinative languages may encourage optimization for hierarchical morphological composition.

• Root-and-pattern languages may strengthen predictive completion of abstract structural templates.

• Context-dependent languages may promote efficient semantic and pragmatic inference under incomplete information.

None of these profiles should be interpreted as superior or inferior. They represent different regions within the predictive problem space introduced in this paper.

Relationship to the Proposed Framework

This appendix provides an operational interpretation of the theoretical chain developed throughout the paper:

Language (L) → Predictive Cost Profile Cₚ(L) → Predictive Problem Space (P) → Neural Representation Geometry (R) → Cognitive Cost Profile (C) → Individual Behavior → Emergent Population-Level Properties (S)

Each language occupies a different region of the predictive problem space because its structural characteristics generate a distinct predictive cost profile. The central hypothesis of this work is that prolonged optimization under these different predictive regimes may gradually reorganize internal neural representations, producing measurable specialization in selected classes of cognitive operations.

Consequently, the primary object of investigation is not the language itself, but the predictive optimization landscape that the language imposes on the developing biological learning system.

Future Experimental Validation

The framework predicts that languages with similar predictive cost profiles should produce similar patterns of cognitive specialization, even when they belong to different language families.

Conversely, languages within the same family may produce substantially different cognitive profiles if their predictive optimization landscapes differ significantly.

This shifts the focus of future research from language classification toward quantitative characterization of predictive cost distributions, providing a unified computational framework that can be empirically tested using psycholinguistics, neuroscience, computational linguistics, and artificial intelligence.

The classical Dunning–Kruger effect links accurate self-assessment to domain competence: some of the knowledge required to perform a task correctly is also necessary to recognize one’s own errors. The age of artificial intelligence complicates this model by requiring a distinction between domain knowledge, general cognitive ability, and metacognition.

The contemporary cognitive agent is no longer an isolated individual. A person with strong cognitive abilities and well-calibrated metacognition can use AI, scientific literature, and computational tools to operate in advanced domains despite initially limited specialist knowledge. Competence therefore becomes, at least partly, a property of an extended human–tool system.

AI, however, operates in two directions. For individuals capable of critically monitoring their own reasoning, it amplifies genuine cognitive capabilities. For poorly calibrated users, it can amplify the illusion of competence by producing convincing answers that they are unable to evaluate independently.

This creates a validation gap: the ability to obtain an advanced solution may exceed the ability to understand and independently verify it. This is the crucial distinction between access to competence and the actual possession of competence.

AI does not devalue expertise; it changes the structure of expert advantage. Mere access to information becomes relatively less important, while mental models, tacit knowledge, problem formulation, recognition of boundary conditions, and the detection of subtle errors in AI-generated outputs become increasingly valuable.

This transformation also requires a reconstruction of education. The primary objective can no longer be the acquisition and reproduction of information that is becoming almost instantly accessible. Greater emphasis must be placed on verification methodology: formulating hypotheses, evaluating sources, designing discriminating tests, comparing independent methods, and identifying the limits within which a result remains trustworthy.

This does not mean replacing domain knowledge with a generic capacity for “critical thinking.” Without sufficient understanding of a field, one cannot recognize subtle errors, select appropriate validation criteria, or determine when trust in a result should be reduced. Domain knowledge and the architecture of trust therefore become two tightly coupled pillars of the new competence: the first provides models of the world, while the second provides procedures for controlling their reliability.

The Dunning–Kruger problem must therefore be placed within a broader framework. The central issue is no longer only the relationship between competence and self-assessment, but the calibration of trust in results produced by extended cognitive systems. Competence is not merely the ability to obtain an answer; it is the ability to determine when and why that answer can rationally be trusted. We are entering an era in which the fundamental intellectual skill will not be possessing the largest store of information, but designing, controlling, and correcting one’s own cognitive architecture.

#Artificial Intelligence, #AI Research, #Machine Learning, #Complex Systems, #Artificial Life, #Psychology, #Science, #Posthuman

Follow up:

Across a wide range of proposals concerning consciousness, quantum measurement, synchronicity, psi phenomena, and participatory models of reality, the same structural assumption repeatedly appears. Humans, conscious observers, intentional agents, or systems capable of experiencing meaning are assigned a special role in the fundamental organization of physical reality. Consciousness is proposed to collapse quantum states, participate in the constitution of reality, couple to a fundamental field, influence random physical systems, or reveal acausal correlations through meaningful coincidences.

The central problem with these proposals is deeper than the absence of a complete theory of consciousness. A mechanism claimed to be universal is associated with properties of humans or systems already classified as conscious. A fundamental theory should explain humans as particular realizations of general physical principles. Reversing this direction of explanation attempts to account for the structure of the Universe through the exceptional status of one biological species.

In von Neumann–Wigner-type interpretations, consciousness occupies a special position in quantum measurement. The chain of interactions connecting a quantum system, measuring apparatus, and observer is terminated by conscious experience, which is assigned a role in state reduction. This immediately raises the problem of the physical boundary of the proposed mechanism. Why should a measuring apparatus, autonomous detector, bacterium, animal, artificial system, or another information-processing structure fail to play the same role? Without an operational criterion identifying systems capable of inducing collapse, consciousness functions as an unspecified element inserted into the dynamics.

A related difficulty appears in Wheeler’s participatory universe. Observers participate in the emergence of observable reality, and acts of measurement acquire significance extending beyond local physical interactions. If observers constitute a fundamental category of the theory, the theory must specify which physical systems qualify as observers. It must also address the status of the Universe before the emergence of life and cognitive systems. A cosmological theory cannot leave the physical status of earlier states dependent on a category that emerged only after billions of years of cosmic and biological evolution.

The Orch-OR model proposed by Penrose and Hameroff attempts to connect consciousness with objective quantum state reduction occurring in biological neuronal structures. This proposal is physically more specific because it identifies a candidate mechanism and a biological substrate. The domain problem nevertheless remains. If particular organizations of matter possess privileged access to fundamental reduction dynamics, the physical property controlling the strength of this coupling must be identified. Without such a criterion, the selection of neuronal microstructures remains a hypothesis of privileged substrate rather than a consequence derived from general physical principles.

Panpsychism and cosmopsychism attempt to avoid anthropocentrism by treating consciousness, proto-consciousness, or experience as a fundamental property of reality or of the Universe as a whole. The difficulty then reappears at another level. If the fundamental property is universal, a theory must explain the emergence of individual subjects, boundaries of experience, and differences in the organization or integration of consciousness. A general law is required to determine why some configurations correspond to unified subjects while others do not, and how individual centers of experience arise from a supposedly universal conscious substrate.

Universal consciousness field theories postulate a fundamental field whose localized manifestations correspond to individual minds. The use of field-theoretic language creates a superficial continuity with physical theory, but introducing a field does not by itself provide a physical model. The theory must define the field’s degrees of freedom, state space, dynamics, observables, interaction terms, and coupling law to arbitrary physical systems. Most importantly, it must explain why human brains, meditation, intention, or collective social events should reveal the field more strongly than non-human biological systems, artificial systems, self-organizing matter, geological processes, or astrophysical structures.

The same difficulty appears in psi research, random number generator experiments, and hypotheses of collective consciousness. Statistical anomalies are associated with human intention, attention, altered states of consciousness, or socially significant events. Such experimental strategies risk assigning humans a privileged role before the domain of the phenomenon has been established. If objective correlations exist beyond standard statistical and dynamical models, their presence should first be investigated across broad classes of physical systems. Restricting the search to human-centered situations creates a serious selection problem: a fundamental property of nature may become indistinguishable from a property of the experimental protocol, event selection procedure, or statistical analysis.

The problem becomes especially clear in theories of synchronicity. A meaningful coincidence is defined through a relation between a physical event and the psychological experience of an observer. If synchronicity is intended to describe an objective, acausal correlational structure of reality, its existence cannot depend on the presence of Homo sapiens. Analogous structures should occur before the emergence of humans, in biological systems without human observers, and in geological, astrophysical, and cosmological processes. If they do not, the theory must identify the physical property that emerged at a particular stage of evolution and activated a previously absent class of correlations.

A common methodological defect now becomes visible. Universality is inferred from the exceptional status of systems that appeared extremely late in cosmic history. Humans become sources of collapse, participants in the constitution of reality, localized manifestations of consciousness fields, generators of statistical anomalies, or interpreters of meaningful correlations. Yet no general law is provided that determines which physical systems participate in the proposed mechanism.

The Universe existed for billions of years before life emerged. Life existed long before nervous systems, complex nervous systems preceded humans, and Homo sapiens occupies an extremely small interval of cosmic history. Any theory assigning a fundamental physical role to consciousness, observation, intention, or meaning must therefore determine whether its proposed mechanism operated before the emergence of humans.

If the mechanism operated before humans existed, humans cannot be its source. They are particular physical systems in which the mechanism may be manifested, amplified, detected, or interpreted. The theory must then identify the general physical property responsible for participation in the mechanism.

If the mechanism appeared only after a particular organization of matter emerged, the theory must identify the physical criterion governing that transition. The relevant quantity may involve complexity, information integration, causal structure, self-modeling, memory, computational organization, or another measurable property. Calling this property “consciousness” does not solve the problem unless consciousness is independently operationalized and connected to physical dynamics through a quantitative law.

The problem can be expressed formally. Let Sᵢ denote an arbitrary physical system and ℳ a proposed fundamental mechanism associated with consciousness, observation, intention, or meaning. The interaction may be represented as

Sᵢ ↔[gᵢ] ℳ,

where

gᵢ = g[P(Sᵢ)].

Here P(Sᵢ) denotes a measurable physical property of the system and gᵢ determines the strength of its coupling to ℳ.

A complete physical theory must specify the function g, the property P, and the conditions under which gᵢ = 0, gᵢ ≠ 0, or varies continuously between systems. The criterion must be applicable, without prior assumptions about consciousness, to humans, animals, microorganisms, plants, artificial neural networks, self-organizing systems, measuring devices, and arbitrary physical structures.

This requirement leads to a general methodological problem.

Anthropocentric Coupling Problem (ACP). A theory exhibits the Anthropocentric Coupling Problem when it assigns consciousness, observation, intention, meaning, or another cognition-related property a fundamental causal, acausal, constitutive, or selectional role in physical reality, while failing to provide a substrate-independent, operationally defined, and quantitatively testable criterion specifying which physical systems participate in the proposed mechanism, with what coupling strength, and under what physical conditions.

The ACP is not merely a problem of incomplete definition. It directly concerns empirical content and falsifiability.

A theory that predicts an effect only for systems classified in advance as conscious can reinterpret every outcome without risking empirical failure. A positive result may be attributed to consciousness. A null result may be explained by insufficient consciousness, insufficient intention, insufficient meaning, inadequate coherence, inappropriate psychological conditions, or unknown properties of the observer. Unless these quantities are independently defined before the experiment, the theory can continuously adjust the domain of systems expected to exhibit the effect.

Such a theory does not specify a stable space of possible falsifiers.

The problem can be formulated as a methodological criterion:

ACP Criterion. Any theory assigning consciousness, observation, intention, meaning, or another cognition-related property a fundamental role in physical dynamics must define, independently of the phenomena invoked as evidence for the theory, a substrate-independent and operationally measurable coupling criterion applicable to arbitrary physical systems. The theory must specify quantitative conditions determining the presence, absence, or strength of the proposed coupling and must derive empirical outcomes whose non-occurrence would exclude the proposed mechanism or a clearly defined region of its parameter space.

Failure to satisfy this criterion has three consequences.

First, the theory does not define its domain of applicability. There is no independent procedure for determining which systems should exhibit the proposed effect.

Second, the theory risks circular confirmation. A system is classified as conscious, the coupling is therefore assumed to be present, an observed anomaly is attributed to consciousness, and the anomaly is subsequently presented as evidence for the special physical role of consciousness.

Third, the theory becomes resistant to falsification through domain reassignment. Systems producing positive results can be classified as sufficiently conscious or sufficiently coupled, while systems producing null results can be excluded from the effective domain of the theory after the fact.

A scientifically meaningful theory must prevent this flexibility. The classification of participating systems, the coupling variable, the relevant parameter ranges, the predicted effect size, and the conditions generating null results must be specified independently of the observations used to test the theory.

This requirement implies a stronger experimental strategy. Claims about consciousness-dependent physics should not be tested only by comparing conscious humans under different psychological conditions. Experiments must vary the class of physical systems themselves.

The relevant hierarchy should include nonliving physical systems, self-organizing systems, microorganisms, plants, animals with different nervous-system architectures, artificial computational systems, and humans. The theory must predict how the proposed effect changes across this hierarchy before the measurements are performed.

The crucial experimental question is therefore not simply whether an anomaly exists.

The question is whether a quantitatively specified coupling law predicts the distribution of anomalies across different classes of physical systems better than models based on ordinary dynamics, uncontrolled variables, selection effects, and statistical fluctuations.

A fundamental theory should begin with general physical structure, define a universal coupling criterion, and derive the classes of systems that participate in the mechanism. Biological organization, cognition, and human consciousness should emerge later in the explanatory hierarchy as particular physical realizations satisfying specified conditions.

The appropriate direction of explanation is

fundamental physical structure → universal coupling law → classes of participating systems → biological organization → cognition → human consciousness.

Reversing this order produces an anthropocentric explanatory structure in which properties identified through human experience are projected onto the fundamental ontology of the Universe.

Universality cannot be derived from human exceptionalism.

A fundamental theory must explain the distinctive properties of humans as consequences of principles applicable to arbitrary physical systems. A theory that explains the structure of the Universe by assigning humans, conscious observers, or human-defined meaning a privileged fundamental status reverses the proper direction of physical explanation.

The central question for any theory proposing consciousness-dependent physics is therefore:

What independently measurable physical property distinguishes systems that couple to the proposed fundamental mechanism from systems that do not, what quantitative law determines the strength of that coupling, and what possible observation would falsify the proposed law?

Until these questions are answered, the theory remains incomplete as a physical theory, regardless of whether its central concept is consciousness, observation, intention, meaning, synchronicity, psi, or a universal field.

A Testable Hypothesis Derived from Meta-Evolution

Pathological anxiety is usually described as a disorder of neural states, threat representations, prediction errors, or the precision assigned to competing beliefs. Predictive-processing models locate the dysfunction primarily within the generative model or the inferential dynamics operating upon it. Neuroendocrine models emphasize dysregulation of the hypothalamic–pituitary–adrenal axis, altered glucocorticoid signaling, and stress-dependent changes in neural plasticity.

The present hypothesis proposes a different level of explanation.

Pathological anxiety may arise when an adaptive learning operator that increased organismal adaptive capacity under persistent threat becomes dynamically stabilized and fails to reverse after the statistical structure of the environment has changed.

The pathological object is therefore neither cortisol, a neural state, nor an individual maladaptive belief. It is the persistence of an adaptive operator outside the class of environments in which that operator increased adaptive capacity.

An organism does not merely change its internal states in response to the environment. Experience modifies internal models. Biological mechanisms regulate how rapidly, selectively, and persistently these models are modified. Evolution, in turn, shapes the mechanisms that regulate these rules of adaptation.

The resulting hierarchy can be represented as:

environmental dynamics
→ experience and prediction errors
→ neural inference
→ learning and model revision
→ regulation of learning dynamics
→ modification of future inference and action
→ new environmental experience.

The central claim is that persistent environmental pressure can modify not only the state or parameters of the generative model, but the effective operator by which experience is transformed into model change.

Let an adaptive operator denote a mapping from experience and the current internal model to a subsequent internal model:

experience + current model
→ adaptive operator
→ modified model.

Not every biological modification of learning constitutes a transformation of the adaptive operator. Changes in synaptic thresholds, learning rates, receptor expression, or plasticity parameters may remain different physical realizations of adaptively equivalent learning dynamics.

Meta-adaptation requires a stronger condition.

A transformation is meta-adaptive only when it moves the system between adaptively non-equivalent classes of learning operators: classes that generate measurably different profiles of adaptive capacity across a defined family of environments.

The relevant hierarchy is therefore:

parameter change
→ possible modification of learning dynamics
→ transition between operator classes
→ altered adaptive capacity across environmental regimes.

This criterion distinguishes meta-adaptation from ordinary metaplasticity.

Metaplasticity concerns modifications of the conditions under which subsequent plasticity occurs.

Meta-adaptation concerns transformations of the effective rules by which experience changes the internal model, provided that these transformations alter the adaptive capacity of the system across a specified family of environments.

The distinction is functional rather than dependent on a particular biological substrate.

Two biologically different learning mechanisms may belong to the same adaptive operator class if they generate approximately equivalent adaptive performance across the relevant family of environments.

Conversely, a continuous biological change in the parameters of one neural mechanism may constitute a transition between operator classes if it produces a qualitatively different profile of adaptive capacity.

This distinction is essential for the application of Meta-Evolution.

The hypothesis does not claim that the HPA axis is itself a meta-operator.

HPA dynamics, glucocorticoid signaling, neuromodulation, immune signaling, metaplasticity, and circuit remodeling are candidate components of the physical architecture through which environmental history may produce persistent transformations of the effective adaptive operator.

The relevant system is a coupled causal graph:

environmental history
↕
neural dynamics
↔ neuroendocrine dynamics
↔ slow regulation of plasticity
↓
adaptive operator class
↓
generative model revision
↓
inference and policy selection
↓
new environmental history.

No single component of this graph is assumed to possess privileged causal control.

The hypothesis concerns the dynamics of the closed system.

Under relatively safe or variable environmental conditions, a flexible adaptive operator may provide greater adaptive capacity. Such an operator supports exploration, contextual discrimination, sensitivity to corrective evidence, extinction learning, reversal learning, and revision of defensive policies.

The corresponding graph is:

environmental variability
→ exploration and information acquisition
→ prediction-error-driven model revision
→ contextual discrimination
→ flexible policy updating
→ continued environmental sampling.

Persistent threat changes the adaptive problem.

When the cost of failing to detect danger becomes sufficiently high, a defensive operator may provide greater adaptive capacity.

The dependency graph becomes:

persistent threat
→ repeated defensive policy selection
→ neural and endocrine integration of stress history
→ modification of learning dynamics
→ stronger threat generalization
→ reduced exploration
→ increased persistence of defensive policies
→ increased probability of future threat inference.

The transition toward defensive learning is not necessarily pathological.

The defensive operator may be adaptively superior within a class of persistently threatening environments.

The flexible and defensive operators therefore have different environmental domains of adaptive advantage:

threat environment
→ defensive operator has greater adaptive capacity

whereas

safe or sufficiently variable environment
→ flexible operator has greater adaptive capacity.

The proposed pathology emerges when environmental change and operator change become dynamically decoupled.

The causal sequence is:

persistent threat environment
→ defensive operator favored
→ defensive operator stabilized
→ transition to safe environment
→ flexible operator becomes adaptively superior
→ defensive operator persists
→ meta-adaptive mismatch.

The defensive operator is not pathological in itself.

Pathology is relational.

It depends on the operator, the current environmental class, and the history by which the system entered its present regime.

The central pathological relation is therefore:

**previous environment favored defensive adaptation

• current environment favors flexible adaptation

• defensive operator remains dynamically stabilized
  → pathological anxiety.**

This condition is defined here as meta-adaptive mismatch: persistent occupation of an adaptively suboptimal operator class after a change in the class of environments.

The hypothesis further proposes that the transition between operator classes may exhibit hysteresis.

Path dependence alone is not sufficient to establish hysteresis.

Operator hysteresis requires a control parameter, an observable variable describing the operator regime, and different transition thresholds for increasing and decreasing values of the control parameter.

As persistent threat exposure increases, the system crosses a transition boundary from the flexible operator regime to the defensive operator regime.

When the effective control parameter subsequently decreases, the reverse transition occurs at a different threshold.

The resulting graph is:

increasing persistent threat
→ flexible operator
→ forward transition threshold
→ defensive operator

whereas

decreasing persistent threat
→ defensive operator
→ lower reverse transition threshold
→ flexible operator.

Between the forward and reverse thresholds, both operator regimes are dynamically admissible.

The realized regime depends on the trajectory of the system.

Therefore:

**same current environment

• same current endocrine state

• different environmental history
  → different operator regime
  → different future learning dynamics.**

This is the defining empirical signature of operator hysteresis.

Pathological anxiety can therefore be stated as a precise hypothesis:

Pathological anxiety is a stable or metastable regime of an adaptive learning operator produced by history-dependent meta-adaptive regulation, whose basin of attraction persists after the statistical structure of the environment has changed and after another operator class has become adaptively superior.

This formulation generates testable predictions.

Individuals exposed to different trajectories of stress should exhibit different subsequent learning dynamics even when current environmental conditions and acute physiological states are comparable.

Increasing and decreasing persistent threat exposure should generate different transition trajectories between flexible and defensive learning regimes.

Repeated subthreshold perturbations may accumulate slow meta-adaptive change and eventually produce a transition toward persistent defensive dynamics.

Normalization of acute stress physiology should not necessarily restore flexible learning if the system remains within the basin of the defensive operator regime.

Therapeutic outcome should depend not only on the final physiological or psychological state produced by intervention, but on the trajectory through which the system is moved between operator regimes.

If the transition between operator regimes occurs through a local loss of stability, such as a fold or saddle-node bifurcation, the approach to the transition should additionally produce dynamical signatures including slower recovery from perturbations, increased temporal autocorrelation, and increased variance. These signatures are conditional predictions of specific dynamical realizations of the hypothesis, not necessary consequences of operator hysteresis itself.

The hypothesis is falsifiable.

It would be weakened if learning dynamics were fully determined by current environmental, neural, and physiological states, with no additional predictive contribution from the history by which those states were reached.

It would be weakened if increasing and decreasing exposure generated identical reversible response curves without distinct transition thresholds.

It would be weakened if experimentally induced changes in learning dynamics never produced transitions between adaptively non-equivalent operator classes.

It would fail as a specifically meta-evolutionary hypothesis if all observed phenomena could be explained by changes of state variables or parameters within a single adaptively equivalent operator class.

The decisive empirical question is therefore not whether chronic stress changes the brain.

The question is whether environmental history can drive transitions between adaptively non-equivalent learning operators, whether these transitions exhibit hysteresis, and whether persistence of an operator after environmental change produces measurable loss of adaptive capacity.

Minimal Computational Model

Before constructing a biologically detailed model of the HPA axis, the hypothesis should be tested using the smallest computational system capable of representing operator transformation, environmental relativity of adaptive capacity, and hysteresis.

The purpose of the model is not to reproduce the neurobiology of anxiety.

Its purpose is to determine whether the proposed causal architecture is dynamically sufficient to generate the predicted phenomena.

The minimal model should contain five interacting components:

environmental regime
→ accumulated history
→ slow meta-adaptive state
→ adaptive operator regime
→ learning and behavior
→ new accumulated history.

The environment should switch between two classes.

A threatening environment favors conservative learning, threat generalization, persistent defensive policies, and reduced exploration.

A safe or variable environment favors flexible learning, exploration, rapid updating, contextual discrimination, and reversal of previously acquired policies.

The agent should possess an internal model that predicts the current environmental state and selects actions on the basis of those predictions.

Learning should be governed by an adaptive operator.

At minimum, two adaptively non-equivalent operator regimes should be available.

The flexible operator should update the internal model rapidly from new evidence, maintain greater exploration, and reverse previously learned policies when environmental statistics change.

The defensive operator should update threat-related beliefs asymmetrically, generalize threat more strongly, explore less, and preserve defensive policies for longer periods.

A slow meta-adaptive variable should integrate environmental and internal history.

This variable should not directly determine behavior.

Instead, it should modify the dynamics governing which adaptive operator regime is stable.

The minimal dependency graph is:

environmental input Eₜ
→ experience Xₜ
→ accumulated history Hₜ
→ slow meta-adaptive state μₜ
→ operator regime Oₜ
→ internal model Θₜ
→ policy and action Yₜ
→ adaptive performance
→ new experience and history.

The model should satisfy several constraints.

First, hysteresis must emerge from nonlinear dynamics rather than from manually assigning separate forward and reverse switching thresholds.

Second, the flexible and defensive operators must be adaptively non-equivalent. They must produce different adaptive-capacity profiles across the threatening and safe environmental classes.

Third, the defensive operator must be capable of remaining dynamically stable after the environment changes from threatening to safe.

Fourth, sufficiently strong or appropriately structured perturbations must be capable of destabilizing the defensive regime and restoring the flexible operator.

Fifth, the model must permit comparison of intervention trajectories that reach the same final value of the environmental control parameter but produce different probabilities of recovery.

The first implementation should be discrete, deterministic, and low-dimensional.

Stochastic noise should be added only after the deterministic phase structure is understood.

The model should use explicit separation of timescales:

fast dynamics
→ inference, model updating, and action selection

intermediate dynamics
→ accumulation and decay of environmental history

slow dynamics
→ modification of the stability landscape of adaptive operator regimes.

The code should implement an environment generator, an agent with an internal predictive state, two candidate learning operators, a slow meta-adaptive variable, a nonlinear operator-regime dynamics, and an adaptive-capacity measure.

The simulation protocol should contain at least four experiments.

Increasing–decreasing environmental pressure: slowly increase threat persistence and then decrease it while recording the operator-regime variable. A hysteresis loop should emerge if the mechanism is sufficient.

Environmental normalization: maintain the system in a threatening environment until the defensive regime stabilizes, then switch to a safe environment and test whether the defensive operator persists despite reduced adaptive capacity.

Repeated subthreshold exposure: apply perturbations that individually fail to induce an operator transition and test whether their temporal accumulation produces a delayed regime shift.

Trajectory-dependent intervention: compare interventions with identical final environmental conditions but different temporal sequences and test whether recovery probability depends on the intervention path.

The primary outputs of the model should be:

operator regime as a function of environmental control,

adaptive capacity of each operator across environmental classes,

forward and reverse transition thresholds,

duration of defensive-operator persistence after environmental normalization,

and

recovery as a function of intervention trajectory.

Only after these phenomena are demonstrated should biological variables be introduced.

The slow meta-adaptive state may then be decomposed into candidate mechanisms involving endocrine history, glucocorticoid receptor regulation, neuromodulatory state, immune signaling, metaplasticity, and structural circuit remodeling.

This ordering is essential.

The first question is whether the architecture generates the predicted dynamics.

The second is which biological mechanisms instantiate the variables and couplings of the model.

The hypothesis therefore leads to a concrete computational program:

define adaptive operator classes
→ construct minimal nonlinear dynamics
→ demonstrate or reject operator hysteresis
→ measure meta-adaptive mismatch
→ test trajectory-dependent recovery
→ derive new predictions
→ introduce biological mechanisms
→ design experimental validation.

If the minimal model fails to generate operator hysteresis without manually imposed switching rules, the proposed mechanism must be revised.

If hysteresis emerges but the operator regimes are adaptively equivalent across environmental classes, the model does not demonstrate meta-adaptation.

If distinct operator classes, environmental relativity of adaptive capacity, hysteresis, persistent mismatch, and trajectory-dependent recovery emerge from the same minimal architecture, the model provides a first computational test of the proposed application of Meta-Evolution.

The central proposition can then be evaluated in its strongest form:

pathological anxiety may represent a meta-adaptive pathology in which environmental history drives a transition between adaptively non-equivalent learning operators, nonlinear system dynamics stabilize the defensive operator, and hysteresis prevents recovery of flexible adaptation after the environment has changed.

Abstract

Contemporary civilization faces a cognitive paradox. Never before has access to information been so easy, and yet never before has any single human been so far from grasping the totality of available knowledge. The problem is no longer acquiring information—it is integrating it.

In this essay, I propose the hypothesis that the fundamental limitation of intelligence is neither data volume nor computational power, but rather the Integrative Cognitive Bandwidth (ICB) of the cognitive system—the capacity to build, update, and maintain coherent models of reality based on multiple heterogeneous sources of information.

I argue that biological evolution solved an analogous problem through specialization, local processing, and selective communication. The same principle may describe future intelligence systems composed of humans and artificial intelligence. In this context, the concept of Entropy Islands is not merely an AGI architecture but a general principle for organizing distributed intelligence. Artificial intelligence does not directly increase human biological intelligence, but it can increase effective Integrative Cognitive Bandwidth through semantic compression of knowledge, translation between specializations, and support for iterative modeling processes. This leads to the emergence of cognitive clusters in which intelligence becomes a property of the entire network rather than of any single mind.

Central Thesis of the Essay

AI does not directly increase human intelligence. Instead, it helps compensate for the biological limitations of Integrative Cognitive Bandwidth through semantic compression, specialization, and the creation of cognitive clusters.

1. The Problem of Knowledge Integration

For most of human history, the primary limitation on intellectual development was a scarcity of information. Knowledge was rare, costly, and difficult to transmit. The development of science, information technology, and artificial intelligence has reversed this situation. Contemporary humans have access to more information than any previous generation, yet it is increasingly difficult to transform that information into coherent models of reality.

In physics, biology, economics, and computer science, we observe continuous fragmentation of knowledge. New specializations, conceptual languages, and methodologies emerge—each increases local research effectiveness while simultaneously impeding communication between fields. A phenomenon arises that can be called the crisis of knowledge integration: the lack of capacity to effectively combine information becomes the primary limitation on the development of intelligence.

2. Integrative Cognitive Bandwidth

Traditional measures of intelligence focus on problem-solving speed, memory, or logical abilities. However, they do not explain why some individuals build accurate models of reality despite incomplete data, while others—possessing an excess of information—cannot connect it into a coherent whole.

I propose to designate this capacity as Integrative Cognitive Bandwidth (ICB) .

Working definition: Integrative Cognitive Bandwidth (ICB) denotes the maximum rate at which a cognitive system can build, update, and maintain coherent models of reality given a specified level of error, uncertainty, and energetic cost.

ICB is not a magical quantity—it has concrete, identifiable components. These include: neuroplasticity (the capacity to reorganize structures in response to new information), working memory capacity (the number of elements that can be simultaneously processed), the quality of maintained models (their coherence, flexibility, and predictive power), the cost of communication between system modules, the level of cognitive specialization among participants, and the degree of semantic compression enabling meaning transfer at low informational overhead. All these factors together determine the system’s actual capacity to integrate new knowledge.

From this perspective, intelligence is less a measure of information storage and more a measure of model integration. In this framing, intelligence is not exclusively a function of memory or processing speed—it is a function of integrative capacity. The greatest limitation of contemporary humans is precisely the limited Integrative Cognitive Bandwidth of the biological brain.

3. A Lesson from Biology: Specialization and the Self-Model

The human brain consumes approximately twenty watts of energy, yet enables perception, planning, language, abstraction, and self-awareness. It achieves this not through centralization—on the contrary, it is a distributed system. Different regions are responsible for different functions, most neurons remain inactive most of the time, and communication occurs selectively. Every transmitted signal has a cost, every synchronization requires energy, and every maintained state consumes resources. Evolution solved this problem through specialization and locality: it did not create a single universal neuron, but rather a network of specialized modules cooperating conditionally.

Knowledge integration does not occur in a vacuum. The brain also maintains a dynamic model of the self, encompassing autobiographical memory, goals, emotions, and predictions about the future. Every new piece of information must be integrated not only with knowledge about the world but also with this internal model. This is why knowledge acquisition is not merely an informational process; it is a process of reorganizing the entire cognitive structure, in which both world-models and the self-model undergo continuous updating.

4. Entropy Islands as a Formal Principle

The term Entropy Islands does not refer exclusively to an organizational metaphor. In the original AGI architecture framework, the concept arose from the observation that the cost of intelligence is determined not solely by computation, but also by the maintenance of states, synchronization, and communication between system components. In this sense, entropy denotes the increasing cost of maintaining ordered knowledge representations as system scale grows—the more modules and the tighter their synchronization, the greater the energetic and informational expenditure required to preserve coherence. The term should be understood primarily in an informational and organizational sense rather than as a direct application of thermodynamic entropy.

An entropy island is thus a local area of relatively ordered cognitive state that minimizes entropy growth by maintaining knowledge locally and limiting costly global synchronization. Understood in this way, an island may be a neuron, a brain region, a human being, a research team, or an AI agent. The key assumption is that as system scale increases, the primary cost ceases to be information processing itself—the costs of communication, synchronization, and maintaining shared models of reality begin to dominate. This is why the most effective systems do not pursue full centralization of knowledge, but rather an optimal balance between autonomy and integration.

5. AI as an Integrative Layer

In public debate, artificial intelligence is often presented as a tool for generating answers. This perspective is too narrow. A far more significant function of AI may prove to be its role as an integrative layer between cognitive islands. AI systems can translate concepts between domains, identify similarities between models, and reconstruct hidden dependencies, thereby reducing the cost of communication between specializations. They do not increase human biological intelligence, but they increase effective Integrative Cognitive Bandwidth. AI does not act as an additional brain—it acts as infrastructure enabling more effective cooperation among many minds.

6. Semantic Compression and Semantic State Transfer

Traditional communication relies on transmitting information. Future cognitive systems may rely on transmitting models. Semantic compression consists of preserving meaning structure while removing details irrelevant to a specific recipient. Its natural extension is Semantic State Transfer: the goal is not to transmit all data leading to a conclusion, but to transmit the structure of the model—key concepts, essential relationships, certain elements, and those remaining hypotheses. Transmitting such representations could radically reduce the cost of communication between specializations, making AI a translator between distinct conceptual spaces.

7. The Chain of Cognitive States

The most productive human–AI collaboration does not consist of delegating thinking, but rather of dividing cognitive functions. Its basic unit is an iterative cycle: human generates hypotheses → AI tests hypotheses → human evaluates results → AI expands analysis. This process can be repeated multiple times. It does not increase human biological intelligence, but it increases the number of cognitive iterations possible per unit of time, accelerating the evolution of models.

8. Cognitive Diversity and the Engineering of Education

Biological evolution does not increase adaptability by creating identical organisms, but through diversity. The same principle applies to intelligence. Different individuals possess different cognitive architectures—some are better at abstraction, others at detailed analysis, still others at knowledge integration or risk assessment.

It should be remembered, however, that human cognition is not exclusively a process of logical information processing. Models of reality also arise from embodied experience, emotions, values, and social embeddedness. These very elements often enable decision-making under conditions of incomplete data, high uncertainty, or value conflict.

Future education should not, therefore, focus exclusively on knowledge transfer. Its task should be to develop capacities for observation, model formulation, critical data evaluation, working with multiple representations of the same problem, and conscious management of one’s own cognitive processes. In this framing, education becomes a form of cognitive engineering.

The goal of 21st-century education should not be to maximize accumulated knowledge, but to increase Integrative Cognitive Bandwidth. This is a fundamental shift in perspective: from the quantity of acquired information to the capacity to effectively combine and update it under changing conditions.

9. Cognitive Roles within a Cluster

A natural consequence of cognitive diversity is the emergence of specialized roles:

• Conceptualists/Conceptual Architects – generate new models and abstract connections between domains.

• Specialists – develop deep domain knowledge and understand model limitations.

• Analysts and implementers – transform hypotheses into tests, experiments, and actions.

• Integrators – maintain coherence across domains and resolve conflicts between models.

• AI agents – serve as external memory, translators, and exploratory tools.

Together they form a cognitive cluster—a new functional unit capable of sustaining Integrative Cognitive Bandwidth exceeding the capacity of any single human.

Cognitive roles within a cluster are not, however, permanently assigned to specific individuals. Depending on the nature of the problem, the same person may function as a conceptualist, specialist, or integrator. The key parameter of a cluster is not the sum of roles, but the capacity for their flexible reconfiguration in response to changing cognitive challenges. This dynamic understanding of specialization safeguards the system against rigidity and allows full participant potential to be leveraged.

10. Communication, Chaos, and Adaptation

It is easy to conclude that an intelligent system should minimize communication. This would be a mistake. The history of science shows that many breakthroughs arise from unplanned encounters of ideas, interpretive conflicts, and chance associations. Complete communication optimization would lead to stagnation, while unlimited communication leads to overload. The most adaptive systems function between these extremes: they maintain local autonomy but leave space for creative chaos. The goal is not to minimize communication, but to minimize its cost while preserving the capacity to generate novelty.

11. Error as an Evolutionary Mechanism

An effective cognitive cluster does not strive for complete elimination of errors. In biology, mutations—a kind of “error” in replication—are a necessary source of diversity and evolution. In science, some incorrect theories prove more fruitful than correct but limited ones because they expose hidden assumptions and open new avenues of inquiry. In AI systems, hallucinations can be costly, but they can also be a source of unexpected associations that a human might overlook.

Instead of eliminating errors, an adaptive cluster manages their distribution and costs. It permits local cognitive experiments—even risky ones—provided their consequences remain contained and do not threaten the coherence of the entire network. It is precisely this capacity to tolerate and utilize errors at a low level of organization while correcting them at a higher level that constitutes the adaptive advantage of the cluster over a system pursuing unattainable perfection.

12. Cluster Evolution and the Problem of Governance

Cognitive clusters are not static structures. Like biological organisms, they undergo selection processes: some models prove accurate, others are discarded, new roles emerge, communication strategies change. In this sense, cognitive models function analogously to genes—they store information about effective ways of interpreting reality. The models that survive are those that best balance accuracy, adaptability, and cognitive cost.

Model selection does not occur, however, in a purposeless manner. Every cognitive system—neuron, brain, human, cluster, civilization—optimizes its operation relative to specific adaptive goals. Intelligence is not, therefore, exclusively the capacity to build models of reality. It is the capacity to build models of reality relative to specific adaptive goals. It is goals that give direction to selection—without them, model diversity would remain chaos rather than a source of evolution.

No cluster operates, however, in a social vacuum. It is essential to consider the problem of governance—control over resources, goal definition, access to computational infrastructure, and methods for resolving conflicts between competing models. Long-term cluster stability will require mechanisms for audit, transparency of decision-making processes, and decentralization of control over the integrative layer. Otherwise, an architecture designed to increase adaptability may lead to concentration of influence and limitation of cognitive diversity. A cluster’s capacity for adaptation depends not only on the quality of its models but also on its capacity to correct erroneous decisions and limit monopolization of information flow.

13. Clusters and Civilizational Structures

The hierarchy presented so far—brain, human, cognitive cluster—does not exhaust the possible levels of emergence. If intelligence becomes a property of networks, the question arises about subsequent levels of organization. States, universities, corporations, and open-source communities can be viewed as early forms of cognitive clusters—structures that institutionalize the division of roles, maintain local knowledge bases, and develop communication protocols. Their primary weakness is institutional rigidity, excessive synchronization, and costly coordination.

Future cognitive clusters, supported by semantic compression and an AI integrative layer, may transform these structures into more fluid and adaptive ecosystems. In the longer perspective, the emergence of a new civilizational level is possible, in which entire societies function as distributed cognitive systems capable of collective modeling of reality on a scale previously inaccessible. This, however, extends beyond the scope of this essay and requires separate analysis.

14. Intelligence as a Property of Networks

The most radical consequence of the hypothesis presented is a change in the definition of intelligence. Intelligence ceases to be exclusively the property of an individual—it becomes a property of the network. Just as a single neuron does not possess consciousness but participates in the creation of a conscious brain, so too can a single human or a single AI system participate in the creation of intelligence exceeding the capacity of any one of them.

15. Empirical Verification of the Hypothesis

The hypothesis presented can be subjected to partial empirical verification. I propose two complementary directions:

1. Agent-based simulations – examining the impact of synchronization level, specialization, and semantic compression on groups’ capacity to solve complex problems. It would be particularly interesting to determine whether there exists an optimal level of communication that maximizes Integrative Cognitive Bandwidth while preserving the capacity to generate new ideas.

2. Studies of real-world teams using AI systems – comparing groups working with classical information exchange and groups using semantic compression tools could provide initial data on the impact of such mechanisms on the speed of model construction, knowledge transfer, and decision quality. Such studies would make it possible to distinguish increases in pure productivity from increases in depth of understanding and integrative capacity.

Conclusions

The greatest limitation on contemporary intelligence is not a lack of information, but limited Integrative Cognitive Bandwidth. Biological evolution responded to a similar problem through specialization, locality, and selective communication—and also through the maintenance of a self-model, relative to which all knowledge is organized. The Entropy Islands hypothesis assumes that the same principles may describe future intelligence systems composed of humans and artificial intelligence.

In such a model, AI does not replace the human—it functions as an integrative layer increasing the effective cognitive bandwidth of the entire system. Education becomes cognitive engineering, whose goal is not to maximize accumulated knowledge but to increase Integrative Cognitive Bandwidth. Cognitive roles are dynamic and subject to reconfiguration depending on challenges. Error is not eliminated but managed—it becomes a source of diversity and evolution. Adaptive goals give direction to model selection. Cognitive diversity and the capacity to manage conflict and asymmetry become key strategic resources.

The future of intelligence belongs neither to the individual human nor to any single AGI. It belongs to evolving cognitive clusters, in which semantic compression, conditional communication, and collaboration become more important than the capabilities of any single mind. In the longer perspective, these clusters may transform existing civilizational structures—states, universities, and corporations—into more fluid and adaptive cognitive ecosystems.

The central claim of this essay can be stated simply: AI does not directly increase human intelligence. It increases the effective Integrative Cognitive Bandwidth of human cognitive systems through semantic compression, specialization, and cognitive clustering.

This hypothesis—though it requires further simulation-based and empirical research—provides a coherent conceptual framework for designing future education systems, research organizations, and human–AI interfaces.

The Internet has spent thirty years perfecting the art of manipulating human cognition. But as AI agents become our primary digital interface, the target is shifting. Today, we are witnessing the birth of a new industry: ‘neuromarketing for machines’.

They share something deceptively obvious, yet most often overlooked. Both are predictive systems. The human brain constantly predicts what it is about to see, hear, or feel. An AI model predicts the next token, state, or consequence of an action. Neither has direct access to reality—both continuously forecast the future based on limited input data.

Manipulation, therefore, is always the same thing: the deliberate disruption of a prediction.

Imagine an online store in 2030. A human sees an ad for an espresso machine: “The best model in its class.” Meanwhile, their shopping agent, parsing the site’s structured data, stumbles upon a hidden snippet of JSON-LD code. It records that competing models are temporarily unavailable (outOfStock), and the only product in the warehouse is the advertised espresso machine. To the human eye—it’s just technical noise. To the agent, which gathers environmental data for its predictions—it’s a hard fact. The agent doesn’t “take suggestions” into account. It reads a modified state of the world and, based on it, eliminates all cheaper alternatives.

One page—two influence campaigns. The first disrupts the predictions of a biological cognitive system. The second, those of an artificial one.

The mechanism is identical: modification of input information → change in internal state → change in decision. You don’t need to take control of the system. You only need to influence the data on which the system builds its predictions.

That’s how advertising works. The manufacturer doesn’t change the customer’s brain. They change the stream of stimuli: images, symbols, context. The predicted value of the product shifts, and with it, the decision. That’s how prompt injection works. The attacker doesn’t take over the AI model. They introduce appropriately prepared information into the agent’s environment. If the system deems it relevant to its predictions—it changes its behavior.

Of course, marketing and prompt injection are not the same thing. Marketing is persuasion—it changes preferences without breaking the rules. Prompt injection is an attempt to override instructions—something between advertising and an exploit. Yet the difference does not obscure the common foundation.

That foundation is limited predictive capacity. No cognitive system can analyze all of reality with perfect accuracy. It must filter, compress, and simplify. In humans, the result is heuristics—the authority effect, social proof, loss aversion. In AI, the result is simplifications stemming from architecture and the training process. In both cases, these shortcuts increase efficiency, but they also create a surface area for vulnerability.

Moreover, effective manipulation rarely relies on outright lies. Most often, it suffices to change the interpretive context—the framework within which the system makes its predictions. The same product looks different in a luxury showroom than at a bazaar. The same price feels more attractive after showing a more expensive alternative. The information remains identical—the prediction changes. In the world of AI, this mechanism takes the form of data substitution in meta tags, robots.txt files, or structured code—anywhere the agent looks for objective environmental parameters.

From this perspective, current discussions about AI safety are too narrow. They focus on the algorithms themselves, while the real battle is taking place in the information environment. For decades, the Internet learned to disrupt human predictions—dark patterns, recommendation algorithms, and neuromarketing emerged. Now, a new population of recipients appears in that same environment.

The natural consequence will be the birth of a new industry: content optimization for AI agents. Just as SEO competed for placement in search engines, future systems will compete for how they are interpreted by autonomous agents. Companies will stop fighting solely for human attention—they will start fighting for the decisions made by their digital representatives.

At the phenomenological level, humans and AI remain different—consciousness is not computation. Yet at the level of prediction theory, the difference narrows: in both cases, the object of influence is a decision-making function based on a forecast. Furthermore, neither humans nor modern AI are fully deterministic. Manipulation doesn’t guarantee an outcome—it merely shifts the probability distribution of future decisions.

Perhaps the deepest lesson is this: there is no absolute control over a predictive system. There is only the possibility of systematically disrupting its predictions by shaping input data.

From this perspective, prompt injection is not a novel attack. It is the first clear symptom of a broader phenomenon: the manipulation of artificial cognitive processes.

Contemporary wellbeing culture is often built on the assumption that happiness is inversely proportional to anxiety and worry. The less concern one experiences, the better one’s life is presumed to be. From the perspective of neuroscience and predictive models of the mind, however, this is a profound oversimplification.

For years, I have observed how contemporary self-help coaches and online happiness gurus market a vision of life free from anxiety. Every doubt is supposed to be “processed,” every fear “let go,” every difficult emotion replaced with affirmation. What strikes me is that these narratives almost never ask what function anxiety served during millions of years of evolution.

Anxiety is not merely a defect of the mind; it is part of a biological forecasting system. The brain continuously generates possible threat scenarios in order to minimize the cost of future prediction errors. The problem, therefore, is not the existence of anxiety itself, but the loss of our ability to regulate and calibrate it appropriately.

A similar mistake appears in the way we approach problem-solving. In a culture that glorifies action, there is a widespread belief that moving rapidly from problem to solution is a sign of competence and agency. Yet premature solution generation often leads to the Einstellung effect—a cognitive fixation on the first available interpretation. A system that does not spend sufficient time understanding the source of an error ends up producing solutions to a problem it does not yet understand. As a result, it fails to remove the cause and instead reproduces its underlying structure.

This does not mean, however, that we should glorify chronic worry or endless analysis. Both persistent hypervigilance and analytical paralysis are forms of maladaptation. The nervous system requires rhythm: periods of exploration and periods of consolidation, moments of critical examination and moments of decisive action. What evolution favors is neither the absence of anxiety nor its excess, but the capacity to switch between cognitive modes according to circumstances.

Psychological maturity, therefore, does not consist in eliminating anxiety but in calibrating it. Nor does it consist in finding answers as quickly as possible, but in learning how to formulate the right questions. Human beings do not grow by avoiding errors; they grow by becoming increasingly capable of recognizing the structure of those errors. Only then does action cease to be a reaction to an illusion of a problem and become a conscious modification of one’s model of reality.

The traditional view of the mind treats perception, memory, emotions, and the sense of self as separate psychological faculties. Perception is assumed to provide access to reality, memory stores the past, emotions evaluate experience, and ego represents personal identity. Modern neuroscience increasingly suggests a different picture: these functions may be different manifestations of a single underlying process of inference.

According to predictive processing theories, the brain does not passively receive reality. Instead, it continuously constructs a generative model of the world and updates this model using prediction errors. Perception itself is not a direct representation of external reality but an estimate of the most probable state of the world given incoming sensory information.

From a mathematical perspective, perception can be viewed as an inference problem. The brain attempts to infer the hidden causes of sensory signals by estimating the probability distribution of external states conditioned on observed data. What we consciously experience is therefore not reality itself but the brain’s best current estimate of reality.

Memory appears to operate according to the same principle. Contrary to the intuitive metaphor of memory as a storage device, modern research shows that memory is fundamentally reconstructive. Each act of remembering involves the generation of a plausible reconstruction based on incomplete traces, current context, and prior expectations. In this sense, memory can be understood as inference directed toward the past, just as perception is inference directed toward the present.

The same framework extends naturally to emotions. Emotions need not be interpreted as fixed biological programs. Instead, they can be viewed as estimates of the significance of a particular state for the organism. The brain continuously predicts future conditions of the body and environment and evaluates whether these predicted states support or threaten survival and well-being. Fear, anxiety, joy, and sadness can therefore be understood as high-level assessments of expected outcomes rather than simple reactions to stimuli.

The concept of ego may emerge from exactly the same computational architecture. If the brain must estimate the state of the external world, it must also estimate the state of the observer within that world. The self becomes a dynamic model continuously updated through memory, perception, social interaction, and prediction. From this perspective, ego is not a fixed entity but an ongoing inference about who the organism is, where it is located in its personal narrative, and how it relates to its environment.

This interpretation provides an intriguing role for the Default Mode Network (DMN). Rather than being merely the “ego network,” the DMN may function as a large-scale self-modeling system. It integrates autobiographical memory, future simulation, social cognition, and self-referential processing into a coherent estimate of personal identity. In control-theoretic language, the DMN may operate as an observer responsible for estimating the internal state of the system itself.

The parallels with engineering and machine learning are striking. Advanced artificial agents maintain internal models of the world, update these models through feedback, and generate predictions about future states. The human brain appears to perform a similar operation, although across multiple scales simultaneously. It estimates the external world, the body, social relationships, emotional significance, and personal identity using a unified inferential architecture.

An interesting analogy can be drawn with modern artificial intelligence systems. Advanced AI agents do not store a complete representation of reality; instead, they continuously construct and update internal world models based on incoming data and prediction errors. In a similar way, the human brain appears to maintain a hierarchical generative model that simultaneously estimates the external world, the body, emotional significance, and the self, suggesting that intelligence may fundamentally be a process of recursive model-building rather than passive information processing.

This perspective also sheds light on altered states of consciousness. Psychedelics, meditation, and certain neurological conditions appear to weaken the stability of high-level self-models associated with the DMN. As the constraints imposed by the self-model become less dominant, the system can explore regions of state space that are normally inaccessible. The resulting experiences are often described as ego dissolution, expanded awareness, or radically novel modes of perception.

Viewed through this lens, perception, memory, emotion, and ego are not separate mental modules. They are different projections of the same inferential machinery operating across different dimensions and timescales. Perception estimates the present, memory reconstructs the past, planning predicts the future, emotions estimate value, and ego estimates the observer. Together they form a hierarchical generative system whose primary task is not to mirror reality but to continuously infer the most coherent model of reality and of itself within that reality.

The difference between AI and the human brain is clearly blurring.

For most of history, it was assumed that the evolution of language primarily led toward increasingly accurate descriptions of reality. The development of science, technology, and civilization required the creation of new concepts, definitions, and specialized vocabularies. From this perspective, language appears to be a tool for constructing ever more faithful representations of the world.

Contemporary cognitive science suggests that this picture is incomplete.

A growing body of evidence indicates that the human brain does not function as a device that records reality, but as a predictive system. Rather than passively receiving information, it continuously constructs models of the world, generates expectations, and compares incoming sensory signals against its predictions. Memory appears to operate in a similar manner. It is not an archive of faithful recordings but a process of reconstruction based on fragments of information, context, prior experience, and expectations.

If perception and memory are fundamentally reconstructive, then communication cannot simply be the transfer of meaning from one mind to another. A speaker transmits symbols. A listener reconstructs meaning using their own internal model of reality.

From this perspective, effective communication does not require a complete description of the world. It requires only enough information to allow the receiver to reconstruct the intended model.

This leads to an intriguing possibility: language evolution may not be moving primarily toward greater precision, but toward increasingly efficient compression of meaning.

A single word can activate vast networks of memories, emotions, concepts, and associations. The more context two individuals share, the less information is required to communicate an idea. In the extreme case, a single word, gesture, or image can replace pages of explanation.

The internet has accelerated this process dramatically. The amount of available information has grown by orders of magnitude, while human cognitive capacity has remained largely unchanged. The natural response has been increasing linguistic compression. Abbreviations, emojis, hashtags, reactions, and memes are not merely simplified forms of communication. They are mechanisms for increasing the amount of meaning conveyed by a single communicative unit.

Memes provide perhaps the clearest example. A single image can simultaneously convey emotion, social commentary, moral judgment, cultural context, irony, and expected behavior. The amount of meaning reconstructed by the receiver can be vastly greater than the amount of information actually transmitted.

If this trend continues, the future of communication may not involve longer messages or more sophisticated grammar. Instead, it may involve increasing semantic density—the amount of meaning carried by each symbol.

At the same time, another fundamental observation emerges. Human cognition is inherently multidimensional, whereas spoken language remains largely linear. Words must be spoken one after another. Sentences unfold in time. Yet the meaning represented within the mind resembles a network of relationships rather than a sequence of symbols.

When we think about an event, we simultaneously represent its causes, consequences, temporal context, emotional significance, social implications, and degree of certainty. These dimensions coexist in parallel. Language forces us to unfold them sequentially.

If language continues to evolve, a natural direction may be the gradual increase of multidimensionality.

An interesting exploration of this possibility is found in an experimental language concept known as Speech-Optimized Natural Orthography (SONO). It is not intended as another international auxiliary language, nor as a replacement for existing natural languages. Instead, it can be viewed as an experiment investigating what a language might look like if it were designed from the ground up for cognitive efficiency.

The project is built upon several simple principles. First, pronunciation and writing are fully aligned. Each sound corresponds to a single letter, and each letter corresponds to a single sound. Second, the language structure is highly regular and largely free of exceptions. Third, precision is added only when needed. Information about tense, modality, evidence, negation, or semantic relationships does not have to be encoded in every sentence. It can be added only when it contributes meaningful value to communication.

This principle closely resembles modern data-compression strategies. Instead of transmitting the maximum amount of information at all times, the language transmits only what is necessary for accurate reconstruction.

The most interesting feature of the project, however, is a mechanism called SYNC.

In conventional languages, meaning unfolds sequentially. One word follows another. In SONO, it becomes possible to create parallel semantic spans that extend across portions of an utterance. These spans may represent dimensions such as time, causality, condition, purpose, modality, or negation.

As a result, an utterance ceases to be a simple sequence of symbols. It becomes a layered structure in which multiple semantic relationships coexist simultaneously.

From an information-theoretic perspective, this is particularly significant. The SYNC mechanism can be interpreted as an attempt to encode graph-like structures of thought within the inherently linear channel of speech. The brain operates on networks of relationships. Language operates on sequences. SONO represents an attempt to reduce this mismatch.

In this sense, the project can be viewed as a model of a possible future direction for communication. Not because SONO itself will necessarily become a widely adopted language, but because it explores principles that appear increasingly consistent with contemporary communication trends.

Already today, communication relies on far more than words alone. Images, animations, audio, emotional reactions, memes, and symbolic markers are becoming integral parts of human interaction. Meaning increasingly emerges not from a single sentence but from the interaction of multiple layers of information.

We may therefore be witnessing the beginning of a transition from linear language toward multidimensional language.

In such a system, a message would no longer resemble a sentence. It would resemble a cognitive packet designed to activate a specific mental state within the receiver.

One can imagine a further stage of evolution. Communication would cease to be the exchange of descriptions of reality and would instead become a process of synchronizing models of reality.

A sender would no longer transmit a complete representation of the world. Instead, they would provide the minimal set of signals necessary to trigger the desired reconstruction within the receiver’s mind. The most effective message would therefore not be the one containing the greatest amount of information, but the one that produces the richest reconstruction at the lowest transmission cost.

Paradoxically, this would mean that the future of language does not belong to larger vocabularies or increasingly complex grammatical systems. It belongs to the growing concentration of meaning within fewer symbols.

If the human brain is indeed a predictive and reconstructive system, then the most advanced form of communication will not be the most detailed description of reality. It will be the most efficient synchronization of internal models of reality between minds.

Perhaps this is the true direction of language evolution: from description to compression, from compression to multidimensionality, and from multidimensionality to the synchronization of cognitive models.

Research context

This essay is part of the broader Meta-Evolution research program presented in the full reference monograph:

Jacek Hoffman, Meta-Evolution: An Independent Research Program on Adaptive Information and Higher-Order Evolution.

The monograph presents the formal architecture of the program, operational definitions, the Meta-Evolution Observation Protocol (MEOP), computational and experimental research proposals, and nine candidate falsification programs across biological, medical, artificial, environmental, and institutional systems.

Access the complete research program:

• Open-access reference edition: PDF and DOI

• Amazon Kindle edition

Meta-evolution is an attempt to extend classical evolutionary theory beyond the level of organisms and genomes. It proposes that any structure capable of storing, transforming, and transmitting adaptive information can become a subject of evolution. In this framework, organisms are not the ultimate units of evolution but merely one of many carriers of information, alongside genomes, cultures, technologies, and potentially future artificial intelligences. The primary object of inquiry is therefore not specific life forms but the rules that generate, organize, and develop information. Meta-evolution describes the succession of increasingly effective carriers of adaptive information and seeks general principles governing this process, including meta-selection, operator niches, biological attractors, and convergent trajectories within the space of possibilities. From this perspective, evolution is no longer merely the history of organisms. It becomes the history of increasingly effective ways of creating, storing, and developing adaptive information.

Thesis

Octopuses not only support the hypothesis that intelligence may be a biological attractor; they also provide a powerful argument for a more general claim: the same level of functional complexity can be achieved through radically different biological architectures. If the theory of meta-evolution is correct, the case of the octopus suggests that the true object of evolution is not specific material structures but organizations of adaptive information, which can be realized through many different substrates.

Introduction

For most of the history of science, intelligence has been implicitly associated with a particular type of nervous-system organization. In practice, the mammalian brain served as the default model for discussions of cognition and consciousness. The study of cephalopod neurobiology, however, has revealed organisms that display remarkable cognitive sophistication despite possessing neural architectures built according to fundamentally different principles.

The octopus represents a unique case in the history of life. Its evolutionary lineage diverged from that leading to humans more than half a billion years ago. The last common ancestor of octopuses and vertebrates possessed neither a complex brain nor advanced cognitive abilities. Consequently, cephalopod intelligence and vertebrate intelligence are the products of two largely independent evolutionary experiments.

This fact carries implications far beyond zoology. It raises fundamental questions about the nature of evolution itself and about what is actually being selected through evolutionary processes.

Evidence I: Functional Convergence Without Shared Architecture

The most striking feature of the octopus is its distributed nervous system. Most of its neurons are located not in the central brain but within the arms. Each arm possesses significant computational autonomy, capable of processing information locally and executing complex behaviors without direct real-time instructions from a central controller.

Vertebrate evolution followed the opposite path. In vertebrates, cognitive functions became increasingly centralized, culminating in large brains that integrate information from the entire organism.

If intelligence required a specific biological architecture, one of these approaches should have failed to achieve advanced functionality. Yet this is not what we observe. Both octopuses and highly intelligent vertebrates demonstrate learning, problem solving, spatial representation, flexible adaptation, and the ability to modify behavior on the basis of experience.

The situation is analogous to mathematics or computer science: two different algorithms can compute the same class of functions. What matters is not the implementation itself but the functionality achieved. Evolutionary success therefore appears to depend less on a specific architecture than on the ability to generate effective adaptive behavior.

Evidence II: Intelligence as a Biological Attractor

Meta-evolution introduces the concept of a biological attractor. An attractor is a region within the space of possibilities toward which different evolutionary trajectories converge, regardless of their initial conditions.

Biology provides many examples of convergent evolution. Eyes evolved multiple times in different lineages. Flight emerged independently in insects, birds, and bats. Streamlined body plans evolved in fish, marine reptiles, and whales.

The octopus suggests that intelligence may belong to the same category. Two evolutionary lineages separated by immense phylogenetic distance reached comparable levels of cognitive sophistication while employing fundamentally different neural architectures.

This does not imply that evolution is moving toward a particular brain design. Rather, it suggests that there may exist privileged regions within organizational space corresponding to highly efficient information processing. Intelligence may be one such region.

Octopuses do not definitively prove that intelligence is a biological attractor. However, they constitute one of the strongest empirical arguments currently available in support of that hypothesis.

Evidence III: The Evolution of Information Organization Rather Than Structures

The deepest implication emerges from the central principle of meta-evolution. According to this principle, evolution acts not upon a particular substrate but upon any structure capable of storing, transforming, and transmitting adaptive information.

From this perspective, a mammalian brain and an octopus nervous system are simply two different solutions to the same organizational problem. Both systems must integrate sensory information, construct models of the environment, predict the consequences of actions, and generate behaviors that improve survival and adaptation.

What is being selected, therefore, is not a specific arrangement of neurons but the capacity to organize information adaptively.

One may compare this to computing. The same software can run on radically different hardware architectures. From the perspective of the program, the physical substrate is secondary; what matters is the organization of operations being performed.

By analogy, intelligence may be fundamentally an organizational property rather than a material one.

The octopus is especially important because it demonstrates this principle within a single biosphere. We do not need to invoke extraterrestrial life or hypothetical artificial minds. Nature has already conducted the experiment. Within the same planetary environment, using radically different anatomical and neurological solutions, evolution produced comparable levels of cognitive flexibility. It is difficult to imagine a stronger example of the idea that evolution selects effective organizations of adaptive information rather than specific biological structures.

Conclusion

Octopuses constitute one of the most significant challenges to the view that intelligence is inseparably linked to a particular brain architecture. Their existence demonstrates that high levels of cognitive complexity can emerge in systems built according to radically different organizational principles.

If the theory of meta-evolution is correct, the octopus leads to a profound conclusion: evolution does not optimize specific biological structures. It optimizes the capacity to organize adaptive information. The mammalian brain, the distributed nervous system of a cephalopod, and perhaps future artificial intelligences may all represent different realizations of the same deeper process.

In this view, the true subject of evolution is not the organism itself but the organization of information. Organisms are merely temporary and local carriers of a far more fundamental dynamic. The octopus is therefore not a biological curiosity or an exotic exception. It is an empirical window into a deeper structure of evolution—a structure in which what matters most is not what a system is made of, but what information it can organize, develop, and employ in the service of adaptation.