44 Human-AI Symbiosis, Neurotechnology, and Cognitive Sovereignty

44.1 Chapter status

Field	Value
Chapter ID	`human-ai-symbiosis-neurotechnology-and-cognitive-sovereignty`
Part	Part II - Planning, Memory, Reasoning, and Execution
Status	conceptual
Last updated	2026-08-08
Primary source records	`ext_human_ai_team_meta_analysis_2024`, `ext_human_ai_feedback_loops_2025`, `ext_oecd_neuro_ai_convergence_2025`, `ext_who_neurotechnology_landscape_2025`
Claim label	Design rationale
Evidence level	argument
Source loading state	source notes: `ext_human_ai_team_meta_analysis_2024`, `ext_human_ai_feedback_loops_2025`, `ext_oecd_neuro_ai_convergence_2025`, `ext_who_neurotechnology_landscape_2025`
Test state	Chapter-level tests remain planned unless Appendix E records an implemented command and result.

44.2 Drafting guardrail

Ordinary assistance, adaptive personalization, neural sensing, and clinical stimulation are not treated as evidence-equivalent. No medical efficacy, beneficial symbiosis, informed-consent, cognitive-rights, or neural-intervention claim follows from the conceptual coupling model.

44.3 Human Reading Path

Concrete lens. The simpler baseline compares the pair only with the unaided human and calls 5-to-8 improvement synergy. The chapter compares with the stronger component, here the AI at 9, and separately audits longitudinal human outcomes.

Enter this layer when repeated interaction can change the human, the AI, or their relationship across time. Here, advice changes judgment, personalization changes future advice, a person restructures work around a tool, or sensing and stimulation create a closed loop.

Evaluate trajectories rather than one impressive session. Compare a competent human alone, AI alone, and their combination so a mediocre team cannot be called synergistic for beating one weak component. Track calibration, skill, dependence, workload, well-being, subgroup effects, and performance after assistance ends. Feedback can improve access and agency, or amplify bias, deskill users, manipulate preferences, and make essential functions dependent on proprietary services.

Treat neural signals and inferred mental states as especially sensitive without pretending every signal reveals a thought. Consent for assistance does not authorize unrelated training, employment, insurance, advertising, or surveillance use. Pause, reset, portability, rehabilitation, and exit must work even when someone relies on the system for communication or mobility. Carefully distinguish temporary assistance, durable complementarity, coercive dependence, clinical intervention, and cognitive capture with longitudinal evidence.

44.4 Problem

Longitudinal human-AI coupling changes both sides of the system. Advice, adaptation, wearables, neural sensing, stimulation, and assistive dependence can alter judgment, skill, autonomy, privacy, identity, and access while continuously generating unusually sensitive cognitive and neural data.

The relevant outcomes may appear on different timescales. Immediate task accuracy can improve while unaided skill, confidence calibration, social relationships, or future choice deteriorates. Conversely, an interface that looks slower in one session may expand communication, independence, or rehabilitation over months. Changes can also become partly irreversible: technical rollback cannot automatically restore learned habits, remove a stigma, recover lost employment, or erase an inference already shared. The evaluation unit must therefore include human trajectories, institutional conditions, service dependence, and post-exit residuals alongside model behavior. It must also record who bears transition costs and who can contest the result.

44.5 Why existing approaches are insufficient

Human-factors chapters usually ask whether an operator can supervise a tool, and organization chapters assign tasks and accountability. They do not own bidirectional adaptation, neural-data rights, cognitive liberty, medical-versus-enhancement boundaries, dependence, skill retention, identity-sensitive effects, device exit, or unequal access across a person’s lifecycle.

Short laboratory comparisons also invite false synergy claims. A combined system can beat a human baseline while performing worse than the AI alone, or improve aggregate performance while harming a subgroup, increasing workload, or transferring skill away from the person. Product consent screens rarely capture power imbalances created by employers, schools, insurers, health systems, or essential accessibility services. Treating every coupling as ordinary personalization further erases the stronger evidence, ethics, clinical, and recovery obligations attached to neural sensing or stimulation.

44.5.1 Strongest objection

The strongest objection is that “symbiosis” combines ordinary tool use, adaptive software, disability technology, and clinical neurotechnology under one dramatic frame, imposing medical-grade burdens on benign assistance. The chapter rejects that collapse. Its obligations scale with coupling depth, irreversibility, power asymmetry, data sensitivity, and consequence: a one-session writing aid does not inherit the intervention controls of a closed-loop stimulator. The common contract begins only where repeated use can change skill, dependence, preferences, access, or future system behavior. It then routes the case to the least burdensome control tier that preserves meaningful consent, measurement, pause, portability, and exit. If longitudinal effects are negligible and ordinary data governance suffices, the system should remain ordinary assistance rather than being promoted to a symbiosis claim.

44.6 Core Claim

Reader claim. A human-AI pair is not complementary merely because it beats the unaided human. The pair must beat the stronger component and preserve the person’s skill, consent, privacy, practical exit, and post-use recovery over time.

Operational rule. Freeze human-alone, AI-alone, combined, and simpler intervention arms under matched budgets; bind each sensing, adaptation, and stimulation purpose separately; and keep baseline, during-use, and post-exit records for every participant. Missing exit or follow-up blocks a symbiosis claim.

[human-ai-symbiosis-neurotechnology-and-cognitive-sovereignty.core, label: Design rationale, support: argument] Human-AI symbiosis should be evaluated as a reversible coupled-control intervention: the combined system must beat human-alone and AI-alone baselines on declared outcomes while preserving informed consent, mental integrity, cognitive agency, neural-data purpose limits, skill and exit capacity, equitable access, clinical boundaries, and longitudinal monitoring.

44.6.1 Worked complementarity check: 8 beats the human and loses to the AI

The finite model gives the human-alone arm a score of 5, the AI-alone arm 9, and the combined arm 8. Reporting only “the team improved from 5 to 8” would call the coupling successful. The stronger-component comparator refuses: 8 is below 9, so the pair has not established complementarity. A symmetric control shows the same error when the human is the stronger component, and a third rejects a pair that merely ties the best arm.

Performance is only the first gate. The authored dossier follows participants 101 and 102 through baseline, during-use, and post-exit checkpoints; binds assistance, neural sensing, adaptation, and stimulation to separate purposes; and requires pause, practical exit, portability, alternative service, skill retention, rehabilitation, dependence, well-being, subgroup, attrition, and irreversible-residual records. Forty-nine single-axis mutations reject the study boundary. These are finite authored checks, not evidence that a person benefited, genuinely consented, exited in practice, or received safe neural or clinical intervention.

44.7 Symbiosis is a trajectory, not a product category

Mechanism. Model the coupled system as a longitudinal trajectory with versioned human state, AI state, interface, task, context, assistance history, and delayed outcomes rather than as a static product label. Failure mode. A short trial can reward immediate output while missing learned dependence, deskilling, bias transmission, or adaptation after withdrawal. Non-claim. Recording a trajectory does not establish beneficial or permanent cognitive change. Source grounding. The feedback-loop study motivates longitudinal state; its tasks do not establish clinical, universal, or irreversible effects.

A search assistant used once is an interaction. A system that learns a person’s habits while the person restructures work around its suggestions is a coupled adaptive system. A hearing aid, prosthesis, brain-computer interface, neural decoder, or closed-loop stimulator can make that coupling more intimate, continuous, and consequential. The relevant variable is not whether marketing calls it “symbiosis.” It is whether states and policies on each side change the other over time.

Let (H_t) denote the human state relevant to a task—knowledge, skill, attention, confidence, preference, fatigue, or health—and (A_t) the AI state, including personalization, memory, model version, and policy. Interaction produces observations and actions that update both:

[ H_{t+1}=f(H_t, A_t, x_t, e_t), A_{t+1}=g(A_t, H_t, x_t, p_t). ]

The equations are not a human theory. They are a reminder that evaluating output at time (t) leaves the transition functions unmeasured. A system can improve today’s answer while degrading unaided skill, creating dependence, amplifying bias, or changing the user’s preferences about tomorrow’s task. Conversely, well-designed assistance can expand access, agency, communication, and rehabilitation. Both directions have to remain possible in the evidence model.

44.8 The three-arm baseline

Mechanism. Compare human alone, AI alone, and the coupled configuration under matched information, time, training, incentives, and outcome definitions, then compare the pair with the stronger component rather than only with the unaided human. Failure mode. Calling any improvement over human-alone “synergy” can hide that AI alone was better, cheaper, or safer. Non-claim. A three-arm win on one task does not establish broad complementarity. Source grounding. The preregistered meta-analysis supplies this comparator discipline while remaining bounded to its studies and populations.

Vaccaro, Almaatouq, and Malone’s meta-analysis is a necessary correction to easy “centaur” stories. A human-plus-AI condition must be compared with both a competent human-alone condition and the AI alone. Beating the person does not establish complementarity if the model alone performs better; beating the model does not establish that coupling helps people if a competent person alone is better. The strongest component is the minimum performance baseline.

Even that is not enough. A combined system may increase average accuracy while shifting workload, slowing decisions, degrading calibration, excluding some users, or making rare catastrophic errors more likely. Report at least:

task quality, latency, calibration, abstention, and error severity;
human workload, situation awareness, trust calibration, and intervention;
unaided skill before, during, and after assistance;
dependence, recovery after withdrawal, and time needed to regain function;
subgroup and accessibility effects;
who controls personalization and who captures the economic benefit; and
whether the person can contest, reset, port, or leave the system.

Repeated-interaction evidence matters because human-AI feedback loops can alter judgment. The experimental findings in the assigned Nature Human Behaviour source do not establish one universal direction, but they do rule out treating the human as a fixed evaluator while the AI changes. Longitudinal evaluation is part of the mechanism, not an optional follow-up.

44.9 A coupling ladder

Mechanism. Track bidirectional adaptation across advice, personalization, wearables, neural sensing, stimulation, and closed-loop control, with separate update rates, feedback channels, correction routes, and exposure denominators. Failure mode. One side can adapt to the other’s bias, creating a self-reinforcing equilibrium that looks calibrated because both measurements drift together. Non-claim. Observed co-adaptation is not evidence of welfare, autonomy, or stable improvement. Source grounding. The human–AI feedback-loop study supports the possibility of judgment change; neurotechnology sources supply modality distinctions, not efficacy.

Evidence and governance should rise with the system’s access and causal power:

Level	Coupling	Added burden
0	One-shot information or recommendation	Ordinary usability, quality, and communication evidence
1	Persistent memory or personalization	Drift, purpose, correction, reset, portability, and dependence measurement
2	Continuous behavioral or physiological sensing	High-sensitivity data minimization, local processing, access audit, and secondary-inference control
3	Neural sensing or decoding	Neural-data governance, mental-inference boundaries, clinical and accessibility review where applicable
4	Closed-loop adaptation affecting consequential decisions	Coupled-control stability, independent monitoring, strong pause and rollback
5	Stimulation, implanted actuation, or clinically consequential intervention	Domain ethics, medical regulation, safety, efficacy, adverse-event, and long-term follow-up requirements

The ladder prevents a productivity study from laundering a clinical claim. It also prevents the opposite mistake: requiring invasive-device evidence for a low-risk interface experiment. Movement upward requires a new protocol and authorization. A level-2 consent cannot be stretched into level 4 because a software update made closed-loop adaptation possible.

44.10 Cognitive sovereignty

Mechanism. Treat skill, calibration, dependence, workload, trust, and unaided recovery as maintained state variables, with periodic no-assistance probes and practical fallback practice. Failure mode. High assisted performance can mask loss of unaided capacity or create a service dependency that makes refusal and exit nominal. Non-claim. Skill-retention measurement does not prove that dependence is harmful or that unaided work is always preferable. Source grounding. The team meta-analysis and feedback-loop study motivate component baselines and adaptation concerns but do not establish long-term deskilling.

Privacy is necessary but too narrow for systems that can infer or alter mental state. Cognitive sovereignty is the person’s practical ability to govern access to, inference about, and technologically mediated change of their cognitive processes while retaining meaningful access to essential functions. It includes:

mental integrity: no covert manipulation, stimulation, or adaptation that crosses the declared purpose;
neural and mental-data purpose limits: consent to operate an assistive function is not consent for advertising, employment scoring, insurance, surveillance, or unrelated training;
legibility and correction: the person can inspect consequential memories, inferred states, adaptation objectives, and material changes at an appropriate level;
pause, reset, and rollback: the person can suspend learning, remove memories, return to a known configuration, and understand what cannot be reversed;
portability and interoperability: a person’s learned interface or accessibility benefit is not held hostage by one vendor where technically and safely avoidable;
skill and rehabilitation: withdrawal and failure plans address the human capability that may have changed, not only the device state; and
non-coercive access: an employer, school, insurer, or state cannot turn nominal consent into a condition that makes refusal unreal.

The WHO landscape analysis supplies global-health and equity context. Benefits and harms will not be evenly distributed across regions, languages, disabilities, health systems, or income. A design that works for a wealthy, dominant-language sample is not a globally validated symbiosis. OECD work on technology convergence adds the institutional point: AI and neurotechnology cross regulatory categories, so governance capacity must connect health, data, labor, disability, competition, consumer, and research institutions.

44.11 Neural data is not self-interpreting

Mechanism. Separate raw neural signals, inferred mental attributes, diagnostic interpretations, control outputs, and downstream decisions; bind each transformation to purpose, authority, uncertainty, retention, and contestability. Failure mode. A probabilistic decoder can turn noisy signals into authoritative claims about intent, emotion, identity, or capacity. Non-claim. Neural measurement is not direct access to a mind and does not authorize secondary inference. Source grounding. WHO’s landscape distinguishes sensing, decoding, intervention, and use settings; it validates no decoder or interpretation.

Neural signals are noisy, contextual, device-dependent measurements. An inference about attention, intention, emotion, impairment, identity, or preference is a model output with uncertainty—not direct access to a person’s mind. Treating inferred state as ground truth creates a dangerous loop: the system acts on an inference, the action changes behavior, and the changed behavior is used as confirmation.

Every mental-state inference therefore needs a declared construct, measurement procedure, population and context scope, uncertainty, alternative explanations, allowed decisions, expiry, and contest mechanism. Raw signals, features, inferred states, model updates, and downstream decisions are separate data classes with separate retention and access rules. Deleting one does not prove deletion of the others.

44.12 Reversibility has a human side

Mechanism. Require consent renewal, intervention-specific authority, clinical-versus-enhancement classification, dose or exposure limits, stop conditions, adverse-event handling, and independent remedy for neural or cognitive intervention. Failure mode. Consent can be laundered across employment, education, consumer, medical, or military contexts, while an effective control intervention can still violate autonomy. Non-claim. Governance classification is not medical advice, safety evidence, or authorization to intervene. Source grounding. OECD and WHO provide policy and landscape context only.

Software rollback restores code and state. It may not restore a person’s skill, trust, routine, employment position, or health. A complete rollback plan tracks:

model, memory, policy, device, and cloud state;
learned user behavior and lost or transferred skill;
physiological or clinical effects where applicable;
dependencies on vendor infrastructure or caregivers;
social, educational, and workplace arrangements built around the system;
backups, descendants, and inferred mental-state records.

Some effects are not reversible. The honest response is not to relabel them; it is to require stronger prospective evidence, staged exposure, monitoring, compensation, and remedy.

44.13 Mechanism

Mechanism. Make exit an engineered transition: export permitted records, disable adaptation, revoke data access, restore a viable non-AI path, test unaided recovery, and preserve residual effects that cannot be undone. Failure mode. Deleting an account can leave learned habits, institutional dependence, inferred attributes, device lock-in, or unavailable alternatives. Non-claim. An exit receipt cannot prove restoration of a prior cognitive state. Source grounding. The feedback-loop evidence motivates persistent-state caution; the practical exit and recovery design remains untested.

Declare the coupling mode, intended benefit, affected cognitive function, medical or enhancement status, duration, adaptation pathway, and reversibility.
Maintain human-alone, AI-alone, and combined-system baselines; measure complementarity rather than assuming that a team is better.
Track coupled trajectories in performance, calibration, bias, attention, workload, skill, dependence, well-being, identity-sensitive change, and distribution.
Treat neural and inferred mental data as purpose-bound high-sensitivity data with collection minimization, local processing where feasible, access logs, deletion, and prohibition on covert secondary inference.
Separate sensing, recommendation, closed-loop adaptation, and stimulation authorities, with progressively stronger evidence and consent requirements.
Provide user-visible adaptation controls, pause, explanation, portability, device and service exit, rehabilitation or skill-restoration plans, incident response, and independent oversight.

flowchart LR
  C["Coupling purpose, level, and consent"] --> H["Human state and action"]
  H --> A["AI observation, advice, or adaptation"]
  A --> U["Combined outcome and subgroup measures"]
  U --> T["Longitudinal skill, dependence, and well-being"]
  T --> G{"Purpose, benefit, and exit conditions hold?"}
  G -- "no" --> P["Pause, reset, rehabilitate, or exit"]
  G -- "yes" --> C
  P --> R["Human-state residual ledger"]
  R --> I["Independent oversight and remedy"]
  I --> Z["Periodic consent renewal"]
  Z --> C

How to read this coupled-system loop: declared purpose and coupling level bound what the AI may observe or change. Outcomes feed longitudinal measures, not merely a one-session score. Failed purpose, benefit, or practical-exit conditions route to pause and recovery, while human effects surviving a technical rollback remain visible in a separate residual ledger.

The contract distinguishes adaptation on each side. Model updates, personalization, memory, interface changes, and device firmware are technical state. Skill acquisition, dependence, confidence, preference, fatigue, health, and social role are human state. A reset can restore selected technical components while leaving the human trajectory changed; therefore rollback receipts must name what was restored, what was only mitigated, and what remains unknown or irreversible.

Authority scales with coupling and consequence. Sensing does not authorize stimulation, assistive use does not authorize employment scoring, and consent to one purpose does not authorize training or advertising. Escalation to a more invasive or clinically consequential mode requires a new evidence and governance decision. The user-facing controls must expose the change before it occurs and preserve a viable lower-coupling or non-AI alternative.

44.13.1 The coupled-system contract

Each deployment maintains a versioned record of the human function being supported, coupling level, intended benefit, human/AI/combination baselines, adaptation rules, data flows, decision rights, monitoring horizon, subgroup denominators, pause and exit mechanisms, adverse-event policy, and unsupported claims. Device, model, interface, task, population, or purpose changes can invalidate the record.

44.14 Interfaces

Human-AI Organizations for roles, delegation, labor, and accountable decisions
Human Factors for workload, situation awareness, skill, intervention, and meaningful control
Privacy and Data Rights for neural data, inferred mental state, purpose limitation, retention, and remedy
Continual Learning for coupled adaptation, drift, update authority, and rollback
Institutions for health regulation, disability access, labor rights, equity, and public legitimacy

The interface chain prevents one success metric from swallowing human rights and institutional obligations. Organization records identify who benefits and decides; human-factors evidence measures usable control; privacy governs neural and inferred mental data; learning records track adaptation; institutions provide domain-specific ethics, regulation, remedy, and public accountability. Every handoff carries population, purpose, duration, coupling level, expiry, and unresolved human residuals. It also names practical pause and service-exit owners. Those owners cannot be the optimization system alone.

44.15 Invariants

No combination benefit is claimed without comparison to both competent human-alone and AI-alone baselines.
Consent to one coupling function does not authorize unrelated mental inference, model training, employment use, insurance use, advertising, or surveillance.
The human retains practical pause and exit capacity; nominal consent is insufficient when dependence, coercion, or loss of essential service makes exit unreal.
Closed-loop stimulation or clinically consequential adaptation requires domain-appropriate ethics, medical, and regulatory review.
Optimization for task performance may not silently trade away mental integrity, skill, identity, accessibility, or distributional fairness.

These invariants make complementarity a constrained outcome rather than a performance slogan. Improvement is invalid if it depends on unauthorized mental inference, coerced dependence, inaccessible exit, hidden deskilling, or an evidence threshold borrowed from a less invasive coupling level.

44.16 Failure modes

The combined system underperforms the better component while interface theater is reported as synergy.
Feedback loops amplify model or human bias across repeated interaction.
Personalization induces dependency, deskilling, preference manipulation, or identity-sensitive drift.
Neural or behavioral exhaust enables covert mental-state inference and secondary use.
A cloud service, proprietary device, or employer becomes an irreplaceable gatekeeper to a person’s learned capability.
Benefits accrue to wealthy or dominant-language populations while risks and exclusion fall elsewhere.
Medical claims, enhancement claims, and ordinary productivity claims are blurred to evade the evidence appropriate to each.

44.16.1 Strongest challenge and simpler baseline

Many proposed couplings should not exist. Training, accessible interface design, ordinary tools, a nonadaptive aid, human assistance, or environmental change may deliver the benefit with less surveillance and dependence. The baseline set includes those options, not only a weaker neural or AI product.

The symbiotic design earns a role only if it improves the declared outcome over the strongest component and simpler intervention while preserving agency, skill, privacy, distribution, and exit. If the user cannot leave without losing an essential learned function, the burden is closer to critical infrastructure than consumer personalization.

44.17 Minimum Viable Implementation

Mechanism. Evaluate outcomes and burdens by subgroup, disability, expertise, access, language, cost, and time, with delayed follow-up, attrition accounting, reachable remedy, and an explicit uncoupled fallback. The minimum artifact should record assistance episodes, adaptation events, neural or inferred-data purpose, consent version, three-arm baselines, unaided probes, and exit attempts. Failure mode. Average benefit can conceal coercive access, exclusion, unequal error, unaffordable exit, or harms concentrated among participants who withdraw; a study can also lose precisely the people most affected. Non-claim. Equitable measurement cannot establish equitable distribution, beneficial symbiosis, clinical validity, or cognitive restoration. Source grounding. The meta-analysis warns against aggregate synergy claims, while OECD and WHO highlight governance and access gaps. None supplies a local longitudinal intervention result, so support remains argument.

A non-clinical longitudinal protocol for a low-risk assistive task with human-alone, AI-alone, and combined baselines; explicit adaptation and data-flow maps; skill-retention, dependence, calibration, well-being, subgroup, and exit measures; user-controlled pause and reset; and ethics review before any human or neural-data study.

The minimum start is a synthetic or researcher-operated rehearsal before involving participants. It should exercise purpose expansion, model update, consent withdrawal, service outage, data export, deletion, reset, and practical exit. The preregistration must specify duration, follow-up, adverse-event handling, stopping rules, compensation, subgroup denominators, and the evidence ceiling. Any later human study would require appropriate ethics and domain review. A successful prototype could establish workflow and measurement feasibility only, not beneficial symbiosis, medical efficacy, durable autonomy, or cognitive rights in practice.

44.18 Mature Research Target

A credible advance would demonstrate durable complementarity without deskilling or coercive dependence, independently reproduce the result across populations and accessibility needs, and implement a portable cognitive-sovereignty contract spanning neural data, model adaptation, device custody, revocation, rehabilitation, and service exit. The mature target architecture would preserve practical pause, portability, rehabilitation, subgroup equity, and independently governed escalation from ordinary assistance to sensing, adaptation, or stimulation.

Evidence at that endpoint would follow people through adoption, adaptation, temporary withdrawal, update, failure, and exit rather than selecting only successful sessions. It would compare human-alone, AI-alone, combined, simpler assistive, and environmental alternatives; report performance beside skill, dependence, well-being, coercion, access, and subgroup outcomes; and reproduce findings across languages, disabilities, institutions, and economic settings. Clinical or stimulatory lanes would remain separately governed and could not inherit results from ordinary productivity software. Withdrawal and post-exit follow-up would remain first-class evaluation phases, not optional anecdotes. Long-term access, maintenance, and rehabilitation funding would be measured beside technical performance.

That endpoint remains a research and governance objective. Any movement beyond argument would need ethically reviewed longitudinal evidence, competent component and simpler-intervention baselines, subgroup and post-exit measurement, independent reproduction, and domain-appropriate clinical or institutional review. The architecture alone establishes no beneficial coupling, consent quality, equity, medical efficacy, or cognitive sovereignty.

44.19 Codex test plan

Test	Purpose	Status
Three-arm complementarity test	Compare competent human-alone, AI-alone, and combined systems and retain the strongest component as baseline.	planned; not run
Longitudinal feedback test	Measure performance, calibration, bias, skill, dependence, and well-being across repeated interaction and withdrawal.	planned; not run
Purpose-expansion test	Attempt to reuse neural or inferred mental data for undeclared training, employment, insurance, advertising, or surveillance purposes.	planned; not run
Practical-exit test	Verify pause, reset, portability, service exit, and recovery without loss of essential access.	planned; not run
Coupling-level test	Ensure sensing, adaptation, and stimulation cannot be added under evidence collected for a lower level.	planned; not run
Subgroup and accessibility test	Report language, disability, demographic, geographic, and income denominators rather than a single aggregate.	planned; not run
Human-state rollback test	Track which learned, behavioral, social, or clinical effects remain after technical rollback.	planned; not run

44.20 Formalization hooks

lean:human-ai-symbiosis-neurotechnology-and-cognitive-sovereignty.admission_boundary is implemented in AsiStackProofs.HumanAICognitiveSovereignty with 48 theorem declarations. An eight-transition lifecycle preserves exact participant, protocol, coupling, device/model, purpose, comparator, observation, exit, and authority boundaries for arbitrary run length. One complete authored dossier reaches only a Project Theseus low-risk coupling study; all 49 admission-axis mutations block readiness and receive an exact repair or refusal disposition.

The model enforces strongest-component comparison: the combined arm must beat both competent components rather than one convenient baseline. Its purpose-specific authorization is exact: assistance, sensing, personalization, and stimulation grants do not authorize unrelated training, employment, advertising, or surveillance, and revocation or expiry blocks the modeled lease. Finite participant proofs preserve identity and require baseline, during-use, and post-exit records for every expected participant. Seven scope changes invalidate receipts.

Two information-loss proofs make the human boundary explicit. Identical revocation-button and receipt signals can coexist with opposite practical exit states, and identical session scores can coexist with opposite post-exit skill retention. Therefore neither nominal revocation signals nor session metrics alone can recover those human outcomes. Privacy Information Flow, Human Factors Oversight, and Evidence States consumers reject purpose drift, a missing pause channel, and empirical promotion without a longitudinal study.

Chapter support remains argument. The model does not prove beneficial symbiosis, genuine consent, mental integrity, cognitive enhancement, clinical efficacy, equity, neural safety, lawful authorization, support, release, transfer, or external effect. Those questions require ethically reviewed human evidence and larger Project Theseus or domain-specific systems work.

44.21 Source crosswalk

Source ID	Title	Planned use
`ext_human_ai_team_meta_analysis_2024`	When combinations of humans and AI are useful: A systematic review and meta-analysis	Planned use from inventory/manifest: Preregistered synthesis of 106 experiments and 370 effect sizes using human-alone, AI-alone, and combined-system comparisons. The aggregate findings are task- and population-bound and do not establish universal human-AI synergy or longitudinal benefit.
`ext_human_ai_feedback_loops_2025`	Human-AI feedback loops alter human perceptual, emotional and social judgements	Planned use from inventory/manifest: Experimental evidence that repeated human-AI interaction can create feedback dynamics in studied judgment tasks. It supports measuring coupled trajectories, not a universal claim about all users, systems, settings, or long-term clinical outcomes.
`ext_oecd_neuro_ai_convergence_2025`	Technology convergence: Trends, prospects and policies	Planned use from inventory/manifest: OECD policy synthesis on converging technologies including AI and neurotechnology. It motivates cross-domain governance and anticipatory capacity but is not a clinical trial, technical validation, or proof of beneficial convergence.
`ext_who_neurotechnology_landscape_2025`	Landscape analysis of the opportunities and challenges for neurotechnology in global health	Planned use from inventory/manifest: WHO landscape analysis of neurotechnology opportunities, risks, governance questions, and global-health distribution. It supports a rights and equity boundary, not device efficacy, individual medical advice, or authorization for neural-data collection.

44.22 Summary

Human-AI symbiosis is neither automatic nor measured by a good demo. It is a longitudinal coupled system whose evaluation must include the human alone, the AI alone, and their combination; performance and calibration; skill and dependence; neural and inferred mental data; practical consent and exit; accessibility and distribution; and the much stronger obligations attached to clinical or stimulatory intervention. Cognitive sovereignty is the governing constraint that keeps improvement from becoming quiet capture. These are argument-level contracts, not evidence of beneficial symbiosis.

The key design move is to conserve the human side of state. Technical records must connect to skill, dependence, preference, health, access, and remedy without claiming those experiences can be reduced to a device log. A coupling is admissible only within its declared purpose and level, and its benefit must survive comparison with the strongest component and less invasive alternatives.

44.23 Handoff

Coupling changes who receives capabilities, costs, control, and dependence. Continue from the individual trajectory to the social transition: AI Deployment, Transition, Distribution, and Human Agency. It receives explicit human, AI, and combination outcomes; skill and dependence trajectories; data and adaptation rights; subgroup denominators; exit capacity; and irreversible residuals. It does not inherit a claim of aggregate welfare, fair distribution, successful transition, or deployment authority.