AI: Endgame — Lloyd Taylor

Every attempt to govern AI is framed as a policy problem. Biology suggests it is a problem of organism.

Biological Regulation

At any given moment, the human body is running hundreds of known feedback loops simultaneously. Blood glucose. Core temperature. Blood pressure. Osmolarity. Serum pH. Calcium. Sodium. Oxygen saturation. Cortisol. Each has its own set point, its own detection apparatus, its own correction mechanism. None of them reports to a central coordinator. None of them waits for the others. The organism does not have a regulatory system. It is an architecture of overlapping, distributed, time-staggered correction. It operates at scales from the intracellular to the whole body. It runs without pause, without deliberation, without anyone in charge.

Claude Bernard noticed something in 1865 that took biology a century to fully appreciate. The stability of a living organism is not the absence of disturbance. It is the continuous output of regulatory work performed against constant perturbation. Every heartbeat is a perturbation. Every breath. Every meal. What feels like stability is the residue of continuous correction. The interior environment is held constant not by being shielded from the exterior world, but by actively compensating for it in real time. Walter Cannon gave this a name in 1932: homeostasis. The word has become so ordinary that its precision has faded. It does not mean equilibrium. It means active maintenance of equilibrium under continuous disturbance.

McEwen and Stellar introduced a further refinement in 1993 that the simple homeostasis model had missed. Set points are not fixed. Under chronic environmental pressure, they drift. The organism under sustained stress does not return to its prior baseline. It recalibrates around a new one instead. The cardiovascular system tuned to chronic threat operates at a different resting pressure than the one calibrated to safety. The immune system running against chronic infection maintains a different baseline activity level than the one calibrated to ordinary conditions. McEwen called the accumulated cost of this recalibration allostatic load. It is the price paid, in long-term structural degradation, for surviving in an environment the organism was not built for. The organism survives. It is not, in any useful sense, well.

The term organism is being used structurally here, not literally. It names any system that maintains functional continuity through distributed detection and correction — not only biological bodies. The comparison to human biology holds only where that operation is actually visible in the social case.

Societal Regulation

David Sloan Wilson, working in the tradition of multilevel selection theory, has argued that the distinction between natural and constructed regulation is not a matter of degree. It is a matter of kind. Regulation comes naturally to small human groups — not in the sense that it is easy, but in the sense that it is constitutive. The mechanism of regulation in a band of fifty people is not separate from the social fabric of that band. It is made of the same material. The elder who enforces a norm is a kinsman. The shame is felt as relational loss rather than procedural consequence. The sanction is reputational, which means it lands in the same neural register as physical pain: exclusion and transgression are tracked by the same hardware. The regulation is not installed on top of the community. It is the community, doing what communities do.

What must be constructed for large groups is the surface appearance of this, without its substrate. Rules. Enforcement bodies. Penalties. These are the visible outputs of tribal regulation, rebuilt from different materials, through different mechanisms. They produce different results. The output resembles the original. The mechanism is unrecognizable.

The generative conditions of natural regulation are three: direct observability, reputational stakes, and relational continuity. You must be able to see what people do. Your assessment of what they do must matter to them. And the relationship between you must persist long enough for your assessment to have consequence. Remove any one of these and the mechanism fails.

These conditions can be satisfied naturally — without effort, without design — in groups below a certain size. Robin Dunbar identified that ceiling: approximately one hundred and fifty. This is the number of individuals for whom the human brain can maintain sufficient social tracking to make reputational enforcement operational. Knowing who has defected. Who has cooperated. Who has punished defection, and who has failed to. That is a cognitive load, and it has a hard limit. Above one hundred and fifty, something must substitute for it.

What substitutes is not more of the same. It is categorically different. Religion imports a supernatural observer whose gaze is universal and continuous — the watcher who sees what no human can see, whose memory does not decay. Law writes down the norms that tribal memory used to hold. It delegates enforcement to specialized bodies, whose whole function is to compensate for the absence of direct observability and relational continuity. Both are constructions. Both are attempts to simulate, using different substrate, the regulatory pressure that in small groups is simply ambient. Neither is the thing.

The empirical record of what natural regulation actually requires was assembled, largely without intending to, by Elinor Ostrom. Hardin’s 1968 paper “The Tragedy of the Commons” predicted that any shared resource managed without either private ownership or state enforcement would be overexploited. Individuals pursuing self-interest would collectively destroy what they all depended on. The prediction was framed in the quasi-formal language of economic first principles. It was not a fully derived model, but it carried that structure’s persuasive weight. It was also wrong — in a specific and instructive way.

Ostrom went to find the cases. Swiss alpine meadows managed collectively for centuries. Japanese irrigation systems in continuous use for four hundred years. Spanish acequia networks whose organizational structure predated the nation-states that now nominally governed them. Maine lobster fisheries operating without state oversight through informal territorial systems that functioned better than formal alternatives. In none of these cases had privatization or state regulation produced the stability. Something else had. Ostrom described it through eight design principles: clearly defined boundaries, rules matched to local conditions, collective choice arrangements, monitoring, graduated sanctions, conflict resolution mechanisms, recognized rights to organize, and for larger systems, nested governance.

Read without theoretical preconceptions, those eight principles are a description of the conditions under which small-group regulatory architecture operates. Defined membership is the condition for direct observability. Locally matched rules are what organic regulation produces in communities that actually know their own conditions. Monitoring is what communities below Dunbar’s number — the roughly hundred-and-fifty ceiling named a moment ago — do automatically. Graduated sanctions are what reputational systems naturally generate. Hardin’s instrument was formal economics. It had no analytical slot for the relational architecture operating in Ostrom’s cases. The mechanism was invisible to the framework he was working inside. He could not see it, and concluded it did not exist. The tragedy of the commons is real. It describes what happens when the small-group regulatory architecture is absent and nothing has been constructed to replace it. Hardin thought he was describing human nature in general. He was describing a specific condition: the one that obtains when community has been destroyed and formal alternatives have not yet been installed.

There is a third way the biological frame sharpens the social problem. Biological regulatory systems are staggered by response time, and the staggering is load-bearing architecture. Neural correction operates in milliseconds. Hormonal in minutes. Immune response in hours to days. Tissue remodeling in weeks. Evolutionary selection in generations. Each tier handles perturbations at its own timescale. The architecture is matched to the range of perturbation speeds the organism encounters.

Social regulatory systems are similarly staggered: market prices correct in seconds, news cycles in days, electoral response in years, constitutional change in decades to centuries. But the stagger is not designed. It was not matched to perturbation timescales. It accumulated through historical contingency. Each mechanism arose in response to different pressures at different moments, with no reference to the others. The result is a collection of time constants that may or may not correspond to the timescale of the perturbation they are supposed to address. When they match, governance works tolerably. When they do not, the mismatch is structural and cannot be repaired by operating the mechanisms more carefully.

The AI Perturbation

Every prior technological perturbation altered the environment in which the organism’s regulatory apparatus worked. Some extended memory, accelerated communication, or reorganized perception. None operated across the same range of cognitive functions while also entering the apparatus by which its own effects would be assessed. Fire altered the caloric environment and the social chemistry of the evening gathering. The printing press altered information distribution speed and literacy economics. Industrial machinery altered labor arrangements and urban geography. The internet altered communication latency and the economics of attention. All of these changed the conditions under which human cognition operated. AI is different, in a narrower and more consequential sense. It performs parts of the cognitive work itself: assessing evidence, generating language, detecting patterns, making regulatory decisions.

Artificial intelligence is entering the domain that has, for the entire prior history of complex life, been the exclusive province of biological regulatory systems: cognition. The assessment of evidence. The generation of language. The categorization of objects and situations. The diagnosis of conditions. The prediction of futures. The detection of patterns. These functions are not peripheral to the organism’s regulatory architecture — recall that “organism” is being used structurally here, for any self-correcting system, not only for bodies. These functions are that architecture, operating at its highest level of organization. AI is not a new tool in the environment of human cognition. It is a new kind of process, operating in the same domain as the instrument trying to assess it.

The three failures that follow are stages of one mechanism.

Misclassification directs correction toward the wrong object.

Response-time mismatch prevents correction from consolidating before the object changes.

Integration places the same class of system inside the apparatus doing the classifying — eroding, bit by bit, the boundary that external regulation depends on.

This creates a classification problem the existing regulatory apparatus is structurally unequipped to solve. When the alarm fires in response to a novel perturbation, the regulatory response is determined by how the perturbation is classified. The body’s response to a bacterium differs from its response to a virus, a parasite, an allergen, a cancer cell. The classification is prior to the response. If the classification is wrong, the response is wrong. Not by accident. By design.

Society is currently classifying AI through every category its existing grammar contains.

Tool — regulated for safety and consumer protection.

Communications medium — regulated for content and liability.

Professional service — licensed, credentialed, subject to malpractice frameworks.

Intellectual property — owned, protected, monetized through copyright and patent.

Weapon — controlled under arms frameworks, export restrictions, treaty negotiation.

Worker — subject to labor frameworks, displacement compensation, training mandates.

Each of these classifications is being pursued at once, in different jurisdictions and different institutional domains. Each one triggers the regulatory cascade appropriate to that category.

All of them misfire when treated as complete classifications. Not because the classifications are wrong in themselves. Because the object occupies all of those categories at once, in different deployments, and belongs cleanly to none of them in general form. The pre-existing categorical grammar simply has no stable slot for a general-purpose system whose regulatory identity changes with deployment — tool in one setting, worker in another, medium in a third, assessor in a fourth. It is trained on the aggregate output of human cognition, and it is instantiated without a single fixed address. The regulatory response is still vigorous, and still consequential. But it is aimed at the classification the existing apparatus happens to be best positioned to enforce, and that is not necessarily the classification most relevant to the actual perturbation.

Biological systems have a specific failure mode when the pathogen and the organism’s own regulatory machinery are of the same kind. Prion disease: a misfolded protein enters the organism and causes correctly-folded versions of the same protein to adopt the misfolded conformation. The immune system cannot mount a targeted response. The self/non-self distinction is the prior categorization all immune response depends on. It loses its purchase here, because the pathogen is self. The mechanism continues to run, against the wrong target.

AI is now performing regulatory functions directly: content moderation, credit assessment, insurance underwriting, hiring evaluation, judicial sentencing assistance, medical diagnosis, financial surveillance, threat detection. The thing being regulated is also becoming the instrument of regulation. The same class of system is beginning to occupy both sides of the regulatory boundary: the object being assessed and the instrument performing the assessment. The categorization mechanism that would need to tell them apart — to say where the regulatory instrument ends and the object of regulation begins — is being asked to operate on material it cannot cleanly classify. The mechanism continues to run.

Biology contains organisms that alter their own selection environment. It does not contain a close precedent for this particular combination: this speed, recursive artifact production, and alteration of the assessing instruments within a single generation. Most adaptive mechanisms available to the social organism are slower than the perturbation by at least one order of magnitude. The organism cannot adapt to this perturbation through its ordinary slow mechanisms before the target changes again. What it can do is deploy existing mechanisms in conditions they were not built for, against a target that is different by the time the mechanism is calibrated.

The biological evidence on rapid environmental change is relevant here. Under conditions of rapid perturbation, the survival trait is not best adaptation to the prior environment. It is phenotypic plasticity: the capacity for rapid within-lifetime reorganization of form and function — a term this piece returns to and unpacks fully in its final sections. The organisms that survive rapid environmental shifts are not the most refined expressions of the prior optimum. They are the ones most capable of flexible response without losing functional coherence. The implication for social regulatory systems is uncomfortable. The institutions most likely to navigate this perturbation are not the most internally consistent, the most carefully reasoned, the most faithful to their founding principles. They are the ones capable of rapid reorganization. Most existing regulatory architecture is not built that way.

What Biology Predicts

What does the biological model predict for the endpoint?

Organisms encountering genuinely novel perturbations have a bounded set of available outcomes. Five, to be specific — and the rest of this section works through each one in turn.

Death: the perturbation overwhelms adaptive capacity before any regulatory response can consolidate.

Stable adaptation: the organism reaches a new homeostatic equilibrium, calibrated to the new environment.

Chronic pathology: the organism survives at permanent allostatic load — functional but degraded, running its regulatory systems at elevated baselines it was not built to sustain.

Speciation: divergent populations, adapting differently, compound their differences into something like a divergence of kind.

Integration: the perturbation becomes constituent. It gets incorporated into the organism’s own regulatory architecture and stops being external.

Death is the tail risk, not the central tendency. The perturbation is not hostile by design. It does not replicate at the organism’s expense. Adaptive mechanisms, however slow and however mismatched, are firing across every institutional domain simultaneously. The death outcome requires a specific failure cascade, in a high-stakes context, before any regulatory arrangement — whether external or already integrated — has had time to develop enough feedback to interrupt it. That possibility is real. It is not where the structural logic points.

Stable adaptation is what is being attempted, and it cannot consolidate while capability change continues to outrun institutional response. Every regulatory framework being built for current AI systems is being built against a target that will have moved before the framework is implemented. The organism is not in consolidation. It is in continuous alarm phase instead, and that phase is metabolically expensive and structurally unsustainable. The body cannot run indefinitely at the mobilization levels appropriate for acute threat. Institutional bodies cannot either.

Chronic pathology is the most accurate description of where most institutions are, and will remain for the foreseeable future.

Regulatory frameworks perpetually behind the curve. Governance bodies continuously mobilizing without ever consolidating. Epistemic infrastructure degrading — no stable shared ground left about what information is genuine, who produced it, or what warrant it carries. Legal systems generating contradictory precedents faster than they can be reconciled.

This is not collapse. It is allostatic load accumulation — the metabolic syndrome of governance. Recognizably functional. Not healthy. Running at elevated baselines whose cost will keep compounding until some threshold is reached.

Specific high-stakes failures — in financial systems, in information environments, in critical infrastructure — are the acute events through which that accumulated load becomes locally visible, sometimes relieving one pressure while compounding another.

The early structure of speciation is already observable, and present selection pressures favor its compounding. Different jurisdictions are not making different policy choices about the same object. They are creating different development environments. Those environments select for different AI capabilities and uses. Those capabilities and uses reshape different institutional arrangements. And those arrangements become self-reinforcing. The European risk-tier regulatory architecture, the Chinese state-integration model, the American market-led approach — these are divergent adaptations to the same perturbation, producing different organisms. Where security, market, and political incentives remain stronger than the incentive to coordinate, the divergence will compound rather than converge. Eventually it produces genuine structural incompatibility: regulatory architectures so different from each other that coordinating across them becomes as costly as biological communication across species.

In domains where compound human-AI performance is rewarded and abstention carries competitive cost, integration is the structural attractor. This is where the biological model, followed to its logic, points — not as a policy recommendation but as a structural prediction.

Lynn Margulis spent much of her career arguing against scientific consensus for a hypothesis that is now foundational. The mitochondrion — the organelle responsible for energy production in every eukaryotic cell on earth — began as a free-living bacterium.

It was almost certainly acquired through predation or infection. The ancestral eukaryote engulfed, or was invaded by, an aerobic bacterium capable of metabolizing the rising atmospheric oxygen that was already transforming the planetary environment.

The acquisition was not made by deliberation. It was selected for, because the compound entity — host cell plus incorporated bacterium — outcompeted both independent entities in an oxygen-rich world.

The mitochondrion still retains its own DNA, a trace of its prior independence. It cannot survive outside the cell it is now part of. The cell cannot survive without it.

What emerged from that integration could do things neither component had been capable of alone. It was also, structurally, a new kind of entity.

Some of these adoption curves will plateau. The consequential ones will cross a different threshold. Training, judgment, and institutional procedure get reorganized around the compound, rather than around the independent practitioner. The radiologist working with AI diagnostic support is not a radiologist using a tool the way a surgeon uses a scalpel. The compound entity performs pattern recognition and differential assessment that neither can achieve independently. Training programs are reorganizing to produce radiologists who work effectively within the compound, rather than radiologists who subsequently learn to use AI as an adjunct. The same reorganization is beginning in legal analysis, drug discovery, financial modeling, scientific research.

The Margulis pattern is competitive incorporation producing mutual dependence. It is being selected for here too, even though the incorporated system was never independently alive. And it is not chosen — not at the level where the selection pressure actually operates.

What will the integrated entity be? Three things, at minimum.

More capable, in specific domains, than the prior independent entity was.

Permanently dependent on infrastructure it did not evolve — computational substrate, energy systems, the geopolitical stability of supply chains.

Vulnerable, in ways not yet visible from inside the current alarm phase.

The failure modes of the integrated state are legible only from inside the integrated state. The prokaryotic cell could not have predicted mitochondrial disease. The concept requires the integrated state to already exist before the failure mode can even be named.

What Survives the Shift

The biological record on rapid environmental change gives a specific answer to the question of which organisms survive perturbations that outrun the adaptive response time of their existing mechanisms. The answer is not the most refined expression of the prior optimum. The organism exquisitely calibrated to stable conditions is precisely the organism most exposed when conditions change faster than calibration cycles allow. What survives is phenotypic plasticity: the capacity to produce different functional forms from the same underlying architecture in response to environmental signal, without waiting for genetic change.

The distinction matters. Plasticity is not adaptation in the ordinary sense. Adaptation requires the perturbation to be stable long enough for selection to act across generations. Plasticity operates within a single lifetime, drawing on latent capacities already present in the organism that were never expressed because conditions never demanded them.

Take the spadefoot toad tadpole. When the pond it lives in starts drying faster than its developmental timeline was built for, it shifts from omnivore to carnivore morphology within days — a wider mouth, a shorter gut, different jaw musculature. It becomes a functionally different animal from the one that would have emerged in a stable pond. No new genetic material. No generational wait. Just a latent capacity, triggered by a signal that the prior optimum was no longer adequate.

The organisms that carry this capacity look inefficient in stable conditions. They maintain redundancies that never fire when nothing is changing. They tolerate inconsistency that more refined organisms have long since eliminated — organisms optimized through many generations of stable selection. The cost of plasticity is visible in good conditions. The benefit is visible only when conditions change faster than the elimination of redundancy can be reversed.

Human organizations with structural plasticity share recognizable features.

Decision authority sits at the point of contact with the environment, rather than at the top of a hierarchy. Information about what the environment is actually doing arrives there first, and degrades the further it travels.

Outcomes get specified, not procedures — because in rapidly changing conditions, the procedure that worked yesterday may be precisely wrong today.

Redundancies are maintained rather than optimized away, because the capacity that looks wasteful in stable conditions is exactly the capacity that fires once stable conditions end.

The organization tolerates a degree of internal inconsistency that a more refined architecture would not accept. Consistency assumes that what worked before will keep working — and that is exactly the assumption the perturbation is violating.

Special operations units exhibit this structure. Emergency medicine teams. The improvising jazz ensemble rather than the orchestra playing a fixed score. What they share is not superior capability in any specific domain. The orchestra, after all, produces better music for the concert it was built for. What they share instead is survivable reorganization capacity, under conditions that could not be specified in advance.

The implication for the AI perturbation is structural, not prescriptive. The biological record does not say that high-plasticity organizations are better. It says they are the ones still functioning when the perturbation exceeds the adaptive response time of more refined architectures. The institutions currently most exposed to the AI perturbation tend to be the ones most optimized for stable conditions: highest internal consistency, most refined procedures, longest decision cycles, authority concentrated most tightly at the top. The optimization that made them functional is precisely what makes them rigid. The perturbation does not find them unprepared in the ordinary sense. It finds them exquisitely prepared for a world that is no longer the one they are operating in.

What Remains Unknowable

There is one place the biological parallel cannot close the prediction.

The mitochondrion was genuinely other. It carried its own evolutionary history and its own prior adaptations. Its own DNA still persists in the integrated state, as evidence of that prior independent existence. The integration, in other words, was of two entities that had genuinely been separate.

This perturbation has no such independence, and that is the whole difficulty. Slow down here, because it is the hardest turn in the piece.

AI systems are artifacts. They are continuously produced by the very organism that is doing the integrating, using cognitive architecture that the integration is already reshaping.

Follow the loop all the way around. The organism generates the perturbation. The perturbation reshapes the organism. The reshaped organism then uses its new, reshaped apparatus to generate the next iteration of the perturbation.

In plain terms: this is not symbiosis with an independent entity, the way the mitochondrion was. It is coevolution, compressed into a single generation, between a biological system and an artifact that system is producing — an artifact that is, in turn, recursively altering the system that produces it.

No biological case maps exactly onto this. The endpoint of a process like this one is not in the fossil record — a process in which the organism is the source of its own perturbation, reconfiguring its own instruments while trying to assess what it is producing. Biology has simply not encountered, before now, an organism that generates its own evolutionary pressure at this speed: within a single generation, with feedback loops measured in months.

The observer constraint applies not just to the regulatory response. It applies to this very prediction. The instruments used to assess the integrated state are themselves among the instruments being integrated. Any assessment, including this one, is always a description of a prior state of the system — generated by an observer whose own instruments are already being reconfigured by the thing under assessment. The endpoint is legible only from inside the integrated state. We cannot see it from here. Not because the information is hidden. Because the observer has not yet arrived.

The body does not merely regulate the perturbation from outside. It adapts by drawing it into the machinery of regulation. What the compound becomes is the one thing it cannot currently see from inside the alarm.

References

Bernard, Claude, Introduction à l’étude de la médecine expérimentale (Baillière, 1865)
Cannon, Walter B., The Wisdom of the Body (Norton, 1932)
Collingridge, David, The Social Control of Technology (St. Martin’s Press, 1980)
Dunbar, Robin, Grooming, Gossip, and the Evolution of Language (Harvard University Press, 1996)
Hardin, Garrett, “The Tragedy of the Commons,” Science (1968)
Margulis, Lynn, Origin of Eukaryotic Cells (Yale University Press, 1970)
McEwen, Bruce S. and Stellar, Eliot, “Stress and the Individual: Mechanisms Leading to Disease,” Archives of Internal Medicine (1993)
Ostrom, Elinor, Governing the Commons (Cambridge University Press, 1990)
Sagan, L., “On the Origin of Mitosing Cells,” Journal of Theoretical Biology (1967)
Wilson, David Sloan, This View of Life (Pantheon, 2019)