What does AI and AGI mean?

This is the second of three posts on what got eliminated to produce the modern vocabulary of “AI” and “AGI.” The first traced the Dartmouth cut of 1956 — McCarthy’s institutional renaming of cybernetics out of cognitive science, and the consequences that compounded over five decades. The fork I want to look at here is older by six years and structurally distinct. It shaped what “intelligence” would mean operationally — what would count as evidence for it — rather than what it would mean ontologically. The two cuts compound; neither subsumes the other. Dartmouth amputated the questions; the Turing test fixed the criterion. You can have either cut without the other; we got both.

Every benchmark in AI inherits one move made in 1950: the criterion for intelligence is behavioral indistinguishability from a human. ImageNet, MMLU, GPQA, ARC, MATH, HumanEval, BIG-Bench, SWE-Bench, the entire evaluation industry of the 2020s — all of them are descendants of a single substitution Alan Turing made in a single paper. The history is worth unraveling to identify what got eliminated to produce that descent.

Turing Test, 1950

Alan Turing’s “Computing Machinery and Intelligence” (Mind, October 1950) opens with a sentence that gets quoted often and parsed rarely: “I propose to consider the question, ‘Can machines think?’” The next move is the famous one. Turing immediately argues the question is “too meaningless to deserve discussion” because the words “machine” and “think” don’t have agreed-upon technical meanings, and proposes to “replace the question by another, which is closely related to it and is expressed in relatively unambiguous words.” The substitute is the imitation game.

The original imitation game was a three-party setup, which most modern restatements oversimplify. A man (A) and a woman (B) are in one room; an interrogator (C) is in another, communicating with both via teleprinter. C’s task is to identify which is the man and which is the woman; A’s task is to mislead C; B’s task is to help C. Turing then asks: what happens if A is replaced with a machine? “Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman?” The structural innovation is that the test was already an imitation test before the machine was introduced — the man imitating the woman is the conceptual baseline, not direct production of intelligent behavior. The machine is being asked to pull off a gender impersonation that even the man might fail at, against an interrogator with a specific verbal-strategic target. What Turing was actually proposing: not “can a machine think?” but “can a machine pass as a kind of agent in a verbal game where passing is itself the only criterion?” The behavioral substitution is structural, not just operational.

Turing then anticipates and addresses nine objections. The list is worth keeping in mind because most modern arguments against the test recur in his own preemptive list: the theological objection (only ensouled beings can think); the “heads in the sand” objection (the consequences would be so dreadful that we hope it isn’t possible); the mathematical objection from Gödel’s incompleteness theorems; the argument from consciousness (which Turing handles by noting it would force solipsism as the only honest position); arguments from various disabilities (“a machine cannot be kind, resourceful, beautiful…”); Lady Lovelace’s objection (a machine can do only what we tell it to); the argument from continuity in the nervous system; the argument from informality of behavior; the argument from extrasensory perception (which Turing took surprisingly seriously, given the parapsychological climate of 1950 British academia — Joseph Banks Rhine’s Duke laboratory results were ambient at the time). Turing dispatched each objection, sometimes glibly. He had already considered the philosophical objections and decided the operational substitution was worth it anyway. He did not think he was solving the question of machine thought; he thought he was making it discussable.

What forced the substitution was the intellectual climate of the surrounding decade. Behaviorism, descended from John B. Watson’s manifesto “Psychology as the Behaviorist Views It” (Psychological Review, 1913), had become the dominant Anglo-American framework in psychology by the 1930s. B.F. Skinner’s The Behavior of Organisms (Appleton-Century, 1938) and the later, more controversial Verbal Behavior (Appleton-Century-Crofts, 1957) systematized the radical-behaviorist position: mental states are not legitimate objects of scientific inquiry because they are not publicly observable; only stimulus-response patterns are. Clark Hull (Principles of Behavior, 1943) and Edward Tolman (with his more cognitivist “purposive behaviorism,” Purposive Behavior in Animals and Men, 1932) competed within the behaviorist umbrella, but the umbrella covered the whole field. In philosophy, the Vienna Circle’s logical positivism — Rudolf Carnap, Moritz Schlick, Otto Neurath, Hans Hahn — had advanced the verifiability principle as the criterion for cognitive meaning: a sentence is meaningful only if it is verifiable in principle by observation. Introspective claims and metaphysical claims about consciousness were declared either meaningless or reducible to behavioral predictions. A.J. Ayer’s Language, Truth and Logic (Gollancz, 1936) brought the positivist program into Anglophone philosophy with maximum compression and maximum confidence. By 1950, asking “what is it like to be a machine?” was, within the dominant Anglo-American consensus, not a question — it was a confused noise.

Inside that constraint, Turing’s move was to substitute a tractable question for an intractable one: not “can machines think?” but “can machines behaviorally imitate a human verbal performer?” The substitution was legitimate in 1950 — there was nowhere else for the question to go that the surrounding philosophy would accept. What happened next wasn’t Turing’s doing. Every benchmark since inherited the operational substitute without inheriting the acknowledgment that it was provisional. Define a test humans can score, optimize for the test. The optimization target became “pass tests built around humans,” which guaranteed the field would chase human-shaped outputs and never ask what a substrate honestly is when not forced to imitate.

The deeper irony is that the consensus that forced Turing’s hand was already crumbling by the late 1950s. Noam Chomsky’s review of Skinner’s Verbal Behavior (Language, 1959) systematically dismantled the behaviorist account of language acquisition, showing that the empirical data — children producing novel grammatical sentences they had never heard — could not be explained by stimulus-response conditioning. The review is one of the most consequential book reviews in twentieth-century intellectual history; it broke behaviorism in linguistics overnight and accelerated its collapse across psychology. Quine’s “Two Dogmas of Empiricism” (Philosophical Review, 1951) had already begun dismantling logical positivism’s foundational distinctions (analytic versus synthetic, reductionism). Hilary Putnam’s “Brains and Behavior” (1963) attacked the behaviorist account of mental states directly, arguing that mental terms like “pain” cannot be defined by behavioral dispositions because two systems with identical behavioral dispositions can have different mental states (and vice versa). The Hixon Symposium at Caltech (September 1948), where John von Neumann, Warren McCulloch, Karl Lashley, and others presented papers on “Cerebral Mechanisms in Behavior,” is now generally treated as the actual launch of cognitive science as a field — and it pre-dated the Dartmouth conference by eight years. Internal mental states were back on the table by the early 1960s.

Yet the operational test outlived the epistemology that required it. The field kept the constraint after the reason for the constraint dissolved. The tragedy: Turing’s move was historically contingent and locally rational; its persistence is neither. The eliminations that follow are the cost of that persistence — of treating a 1950 expedient as if it were a definition.

Introspection as method

Prior to behaviorism’s ascendancy, there was a substantial pre-behaviorist psychology that took as primary data the trained first-person reports of skilled observers. Wilhelm Wundt founded the first experimental psychology laboratory at the University of Leipzig in 1879 — the conventionally cited founding moment of psychology as a discipline distinct from philosophy. Wundt’s program (set out in Grundzüge der physiologischen Psychologie across editions from 1874 onwards) used trained introspectors performing carefully designed perception and reaction-time experiments. The data was first-person; the methodology was experimental and quantitative. Wundt’s American student Edward Bradford Titchener carried the program to Cornell, refining it into “structuralism” — the systematic decomposition of conscious experience into its elemental components (sensations, images, affections), described in An Outline of Psychology (Macmillan, 1896) and the four-volume Experimental Psychology (1901–05). The Würzburg school (Karl Marbe, Otto Külpe, Karl Bühler) ran a parallel program in Germany from the early 1900s, focused on higher cognitive processes — judgment, deliberation, willed thought.

The methodology collapsed in part on its own contradictions. The Würzburg school argued, on the basis of trained introspection, that “imageless thought” existed — moments of cognition with content but no sensory imagery. The Cornell structuralists argued, on the basis of trained introspection, that there was no such thing — every thought decomposed into images, sensations, and affective tone. The two laboratories could not reconcile their findings using the method itself, and behaviorism’s entry move was to point at this irreconcilability as proof that the method was not scientific. Watson’s 1913 manifesto explicitly cited the Würzburg-Cornell impasse as evidence that introspection had to be abandoned in favor of publicly observable behavior. Skinner consolidated the abandonment.

What was lost is the trained skill of careful self-observation as a source of data about what cognition is. William James’s The Principles of Psychology (Henry Holt, 1890) — two volumes that read as accessibly today as when written — contains the most lucid first-person analyses of attention, the stream of consciousness, the sense of self, habit, will, and emotion in the English language, and most modern cognitive science textbooks cite James as a historical figure rather than as a methodological instrument that could be picked up again. Husserl’s phenomenology — Logische Untersuchungen (1900–01), Ideen (1913), the unfinished Crisis of European Sciences (1936) — developed a rigorous method for the first-person analysis of intentional experience, with technical vocabulary (epoché, eidetic reduction, noesis/noema) that has no entry in any cognitive-science textbook in current use. Heidegger’s Sein und Zeit (Niemeyer, 1927) extended the analysis to temporal-existential structure (Dasein, being-in-the-world). Maurice Merleau-Ponty’s Phénoménologie de la perception (Gallimard, 1945) developed the embodied-perception strand. None of this entered mainstream Anglo-American cognitive science until Francisco Varela’s “Neurophenomenology: A Methodological Remedy for the Hard Problem” (Journal of Consciousness Studies, 1996) attempted the bridge — and even Varela’s program remains marginal three decades later.

Modern recovery attempts continue. Pierre Vermersch’s “explicitation interview” method (L’entretien d’explicitation, ESF, 1994) and Claire Petitmengin’s “micro-phenomenology” (the technical method developed in “Describing one’s subjective experience in the second person,” Phenomenology and the Cognitive Sciences, 2006) are systematic protocols for eliciting trained first-person reports under controlled conditions. Their results — fine-grained empirical data on what cognition actually consists of from inside — are slowly being absorbed into a small corner of consciousness research. The deeper problem they cannot solve quickly: trained introspective competence takes years per person to develop, and the field has not been training it. Two generations of cognitive scientists were taught that their own introspective access was contaminating rather than informative. You cannot reverse that just by funding it.

The methodological loss matters for what I work on. The I=E framework treats every elimination operation as something performed from within a perspective — the move that takes a candidate definition of self or intelligence and shows it can be removed while something keeps registering the removal is, structurally, an introspective operation. It cannot be done from a third-person stance; the third-person stance is one of the candidates being eliminated. So the framework treats trained first-person access as a primary instrument, not as contaminating noise. The Turing filter foreclosed exactly this stance, and the foreclosure compounds: when most of cognitive science treats introspection as suspect, the elimination move I am running has to argue for its own methodological standing before it can argue for any of its results. That’s a tax that wouldn’t exist if the cut hadn’t happened.

For modern AI specifically, the consequence is operational. Large language models are trained without any introspective competence whatsoever — they have no procedure for examining what their own activations are doing while they produce text — and their training procedure includes no mechanism to acquire one. They inherit the introspective gap structurally. When we ask “is this model conscious / does this model understand / does this model want anything,” the field has no agreed instrument for asking the question, because the instrument was eliminated seventy-five years ago.

The “what it is like” question

If the criterion established by Turing is behavioral indistinguishability, then whatever the other “system” is like inside is of no methodological relevance. The question was not answered; it was assumed redundant. Thomas Nagel’s “What Is It Like to Be a Bat?” (Philosophical Review, 1974) put the question back onto the philosophical agenda by arguing that subjective character — the qualitative feel of experience — cannot be captured by objective description. Even if you knew everything about bat neurology, echolocation, and behavior, you still wouldn’t know what it is like to be a bat. David Chalmers’ “Facing Up to the Problem of Consciousness” (Journal of Consciousness Studies, 1995) and The Conscious Mind (Oxford, 1996) sharpened the distinction between the “easy problems” of consciousness (explaining behavior, function, access to internal states) and the “hard problem” — explaining why there is anything it is like to undergo any of it. The hard problem was a recovery move forty-five years after Turing had ruled it out of bounds.

The field’s response was to try to do the work the hard problem demands while staying within the operational filter. Bernard Baars’ A Cognitive Theory of Consciousness (Cambridge, 1988) introduced Global Workspace Theory: consciousness is the broadcast of information across a brain-wide workspace, making information accessible to multiple specialized processes. Stanislas Dehaene developed the neural-empirical extension (Consciousness and the Brain, Viking, 2014). Giulio Tononi proposed Integrated Information Theory in a sequence of papers beginning with “An information integration theory of consciousness” (BMC Neuroscience, 2004), with the central claim that consciousness is integrated information (the quantity Φ). David Rosenthal’s higher-order theories (“Two Concepts of Consciousness,” Philosophical Studies, 1986; Consciousness and Mind, Oxford, 2005) hold that a mental state is conscious when it is the object of an appropriate higher-order representation. Each of these had to fight for legitimacy against a field that had structurally decided the question was either solved (functionalism: consciousness is whatever the right functional organization is) or meaningless (eliminativism: the folk concept of consciousness is a confused holdover that future science will dispense with).

My interest in philosophy of mind sits on this wound. The self-eliminating-observer research program I work on (active on this blog) treats the observer as a fixed point of the elimination operation — the residue that survives every attempt to strip it because the operation that does the stripping is what the observer is. It is the I=E framework’s attempt to address Nagel’s question structurally rather than to dismiss it. The elimination move that runs through all my work is a response to the situation Turing’s filter created: I cannot get behind the first-person perspective to test it from outside; the framework concedes this and treats the perspective as constitutive rather than as an obstacle to be eliminated. Whether that move succeeds is what the program is trying to find out.

For modern AI safety and alignment research specifically, the Turing filter still bites. Most working AI safety researchers treat consciousness as not safety-relevant — the argument runs that we should align behavior regardless of whether the system has phenomenal experience, so the question is moot. This argument is intelligible, and it may even be right tactically. But it depends on the Turing-era assumption that what a system is like doesn’t bear on what it can do, which is precisely the assumption recovery work on consciousness has been undermining for fifty years. The position is held more on inherited methodological commitment than on argument.

Non-Western cognitive traditions

Buddhist abhidharma has roughly 2,500 years of systematic analysis of moment-to-moment cognition via trained introspection. The Theravada tradition’s Visuddhimagga (Buddhaghosa, 5th century CE) is a comprehensive systematic treatment of the meditative phenomenology — the jhanas (absorption states), the analysis of the five aggregates (skandhas: form, sensation, perception, mental formations, consciousness), the doctrine of dependent origination (paticca-samuppada), the moment-by-moment arising and passing of cognition (citta-vithi). The Mahayana abhidharma tradition (Asanga’s Abhidharmasamuccaya, fourth century CE) extends and reorganizes the analysis. The skandha analysis is, structurally, exactly the layered elimination move I run in post 5 of this blog (Stripping the Self) — strip away each candidate-for-self and what remains is the operation, not an entity. Buddhism arrived at the conclusion two and a half millennia before I did; the cognitive apparatus is better-developed and more empirically grounded in trained meditators than anything Western cognitive science has assembled.

Classical Indian philosophy beyond Buddhism is no less developed. The Nyaya school (root text: Gautama’s Nyaya Sutras, ~2nd century BCE; commentary tradition extending to Udayana in the 11th century) developed a sophisticated logic of inference, perception, and cognitive error long before the equivalent moves were made in the West — Nyaya’s analysis of the structure of inferential reasoning and its conditions of validity is recognizable to anyone who has worked through Aristotle, but the technical vocabulary is independent. Advaita Vedanta (Shankara, 8th century CE) developed an analysis of the witness — sakshi, the observer who is presupposed by every act of observation but cannot itself be made an object of observation — that pre-figures the structural move in Wittgenstein’s Tractatus 5.62–5.641 (which I cite in post 5) by twelve centuries.

Daoist accounts of spontaneous adaptive action — Zhuangzi (4th century BCE), the Daodejing — describe wu wei and what later neo-Confucian commentators called li (pattern), which together describe something close to the embodied-predictive-processing account modern cognitive science is reconstructing. The neo-Confucian synthesis (Zhu Xi, 12th century; Wang Yangming, 16th century) developed the analysis of moral cognition and the relation between knowledge and action with a precision the Western philosophical tradition didn’t approach until the twentieth century. Islamic philosophy contributes Ibn Sina (Avicenna, 11th century) on intentionality — the relation between a mental state and what it is about — and Mulla Sadra (17th century) on tashkik al-wujud, the gradation of being, which prefigures process-philosophy moves Whitehead would make in 1929.

Modern recovery is happening, slowly, on the academic margins. B. Alan Wallace’s contemplative-science work (The Attention Revolution, Wisdom, 2006; Mind in the Balance, Columbia, 2009) bridges Theravada/Mahayana practice with cognitive neuroscience. Jay Garfield’s translations and analyses of Madhyamaka philosophy (the Mulamadhyamakakarika of Nagarjuna; Garfield’s The Fundamental Wisdom of the Middle Way, Oxford, 1995) make the technical apparatus available in English. Evan Thompson’s Mind in Life (Belknap/Harvard, 2007) is the most rigorous modern integration of phenomenology, embodied cognitive science, and Buddhist thought. The Mind & Life Institute (whose dialogues began in 1987, convened by Francisco Varela, Adam Engle, and the Dalai Lama, with the formal nonprofit incorporated in 1991) has run a sustained series of dialogues between contemplative practitioners and Western cognitive scientists for nearly forty years; its proceedings are an underused archive.

The consequence of the historical exclusion is concrete. A single culture’s mid-20th-century Anglo-American operationalism became the global filter for what counts as a theory of mind. Vast prior art on intelligence, observation, and self-reference was discarded as “not scientific” — meaning not behaviorally testable. The field reinvented weaker versions of what contemplative traditions had already mapped, often without knowing the prior maps existed. For modern AI, the implication is structural: when an LLM is trained predominantly on English-language Western text and the term “general intelligence” is operationalized via benchmarks built within that tradition, the “general” in AGI inherits the filter. There is nothing general about it. It is the specific cognitive shape of a specific intellectual lineage, masquerading as the universal default because the alternative lineages were filtered out before the training data was selected.

Development and ontogeny

If the target of the test is an adult verbal performer, development becomes decoration rather than essence. Jean Piaget’s foundational work — including The Origins of Intelligence in Children (Routledge & Kegan Paul, English 1953) and the synthesis The Psychology of the Child with Bärbel Inhelder (Basic Books, English 1969) — laid out the four-stage model of cognitive development (sensorimotor, preoperational, concrete operational, formal operational) that dominated developmental psychology for forty years. Lev Vygotsky’s Thought and Language (originally 1934, English MIT Press 1962) and the posthumously compiled Mind in Society (Harvard, 1978) offered an alternative framework: cognition develops through internalization of socially mediated interaction in the zone of proximal development (ZPD), the gap between what a child can do alone and what they can do with skilled help. Both Piaget and Vygotsky treated the child as an active constructor of cognition rather than a passive recipient of instruction.

Esther Thelen and Linda Smith’s A Dynamic Systems Approach to the Development of Cognition and Action (MIT Press, 1994) re-framed development in dynamical-systems terms. Their re-analysis of the A-not-B error — the classic Piagetian finding that infants below about 12 months will continue to reach for an object at location A even after seeing it hidden at location B — showed that the error is not, as Piaget claimed, evidence of a deficit in the infant’s representation of object permanence. It is the natural output of a dynamical system in which reaching trajectories accumulate momentum based on previous reaches. Change the experimental parameters (delay, posture, salience of B) and the error vanishes or appears predictably. Cognition is the trajectory of a coupled system, not the contents of a representational store.

Michael Tomasello’s work on the social-cognitive foundations of human development — The Cultural Origins of Human Cognition (Harvard, 1999), Why We Cooperate (MIT, 2009), Becoming Human (Belknap/Harvard, 2019) — argues that what distinguishes human cognition is not raw intelligence but shared intentionality: the species-specific capacity to coordinate attention with others around a shared goal. This is a developmental-evolutionary claim that has no analogue in transformer training. Developmental robotics (Minoru Asada and the field’s foundational papers in the early 2000s; Giulio Sandini’s iCub project; Angelo Cangelosi and Matthew Schlesinger’s Developmental Robotics: From Babies to Robots, MIT, 2015) attempts to build cognitive systems through developmental trajectories rather than terminal training. It is a small field, marginal to mainstream AI.

A historical irony worth noting: Alan Turing’s other 1952 paper, “The Chemical Basis of Morphogenesis” (Philosophical Transactions of the Royal Society B, 1952), was a foundational contribution to developmental biology, mathematically deriving how diffusion-driven instability in chemical reactions could produce stable spatial patterns from initial homogeneity. The reaction-diffusion model is the basis of much of contemporary morphogenesis research, including Michael Levin’s bioelectric-pattern work (referenced in post 4). The same Turing who in 1950 constrained the field to behavioral output produced, two years later, the technical apparatus the field would need for an alternative ontogenetic account. The morphogenesis paper isn’t in the AI canon; it’s in the developmental-biology canon. The intellectual continuity is uncited within the field that took its founding name from him.

The modern AI implication is sharp. Transformers are trained, not developed. There is no ontogenetic trajectory; the model exists at one logical level (the trained-weights file), and the training procedure that produced it sits structurally outside it. The model has no formal access to its own developmental history. Children acquire language through years of embodied, socially-coupled, structured interaction; LLMs acquire it through a single pass through a corpus, with no analogous developmental scaffolding. The two processes produce statistically similar outputs on certain narrow tasks; whether what they produce is the same kind of competence is exactly the question the operational filter forbids us from asking, because the filter only inspects outputs.

Animal minds as continuous with ours

Turing’s test is verbal. The consequence is direct: non-linguistic cognition got interpreted as “interesting biology” rather than as examples of intelligence we should theorize from. Post 4 of this blog (How Much Can We Eliminate from Intelligence Until It Vanishes?) covers the modern animal cognition literature in detail — Alex the African Grey, New Caledonian crows, octopus distributed cognition, cetacean culture, elephants, slime molds — and I won’t repeat that material here. What’s worth adding for the historical narrative is the longer arc.

Charles Darwin’s The Expression of the Emotions in Man and Animals (Murray, 1872) was the first systematic argument for psychological continuity across species — emotions evolved, and they persist in homologous form across the mammalian lineage. The book sold poorly and was ignored by the behaviorist generation that followed. Donald Griffin’s The Question of Animal Awareness: Evolutionary Continuity of Mental Experience (Rockefeller University Press, 1976) is generally credited with reopening the cognitive-ethology field after fifty years of behaviorist suppression; Griffin had previously been famous for discovering bat echolocation, and he used that authority to reintroduce the question of animal mental experience as a respectable scientific topic. His follow-up Animal Minds (Chicago, 1992) consolidated the case.

The cleaner-wrasse mark test is the most striking recent result. Kohda et al., “If a fish can pass the mark test, what are the implications for consciousness and self-awareness testing in animals?” (PLOS Biology 17(2): e3000021, 2019), reported that the cleaner wrasse Labroides dimidiatus — a small reef fish — passes the standard mirror mark test that had previously been considered evidence of self-awareness only in great apes, dolphins, elephants, and a few birds. The paper provoked extensive methodological pushback; Kohda’s lab responded with the more rigorous follow-up “Further evidence for the capacity of mirror self-recognition in cleaner fish and the significance of ecologically relevant marks” (PLOS Biology 20(2): e3001529, 2022, with co-authors including Redouan Bshary and Alex Jordan), which addressed the methodological objections and replicated the result with a brown mark mimicking an ectoparasite. Either the mark test means less than it was taken to mean, or fish are far more cognitively complex than the standard taxonomy of cognition assumed. Either reading damages the human-shaped silhouette the field built on top of Turing’s verbal criterion.

Roger Hanlon’s work on cephalopod camouflage (over thirty years of research at the Marine Biological Laboratory at Woods Hole; representative paper: “Adaptive coloration in young cuttlefish (Sepia officinalis): the morphology and development of body patterns and their relation to behaviour,” Philosophical Transactions of the Royal Society B, 1988) shows octopuses and cuttlefish executing pattern-recognition and pattern-generation in real time at a level of computational complexity the behavioral literature has not absorbed. Each chromatophore is independently controlled. The animal is producing a high-bandwidth, contextually appropriate visual pattern across its entire body surface in milliseconds, almost certainly without anything resembling explicit symbolic processing. Whether to call that “intelligence” is exactly the question the Turing filter forbids us from asking honestly, because the filter only counts verbal performance.

We have no mature theory of what a mind is that treats language as a specialization rather than as the exemplar. Every theory of intelligence downstream of Turing has a human-verbal-adult silhouette even when it denies it. This is a theoretical loss with operational consequences: when LLMs are evaluated against benchmarks of human verbal performance and called “general,” the generalization is a self-confirming artifact of the criterion, not evidence about the underlying capacity.

Ecological and situated cognition

James Gibson’s ecological psychology (The Senses Considered as Perceptual Systems, Houghton Mifflin, 1966; The Ecological Approach to Visual Perception, Houghton Mifflin, 1979) introduced affordances — the action possibilities an environment offers to a particular organism — and direct perception, the claim that the organism does not infer the world from sensory data but picks up information specifying the world directly through perceptual coupling. Gibson’s work was foundational for the later 4E (embodied, embedded, extended, enactive) cognition movement, and like the Umwelt tradition (covered in post 1), it had to fight for every inch against the representational orthodoxy Turing’s framing had locked in.

The behavior-based robotics wave of the late 1980s and early 1990s was the most consequential brief recovery. Rodney Brooks’ “Intelligence Without Representation” (Artificial Intelligence, 1991, with the actual underlying ideas in his 1986 paper “A Robust Layered Control System for a Mobile Robot,” IEEE Journal of Robotics and Automation) introduced the subsumption architecture: cognition does not need a centralized world model; it can emerge from layers of reactive behaviors directly coupled to sensors and actuators. The empirical demonstration was Genghis, a six-legged robot with no central representation that nonetheless walked competently over uneven terrain. Pattie Maes’ work on autonomous agents (Designing Autonomous Agents, MIT Press, 1990) extended the program. For about five years, behavior-based robotics looked like a serious challenger to symbolic AI’s hegemony. Then deep learning (which is descended from Dartmouth’s symbolic ontology even though it doesn’t manipulate symbols) absorbed robotics back into the centralized-world-model paradigm. Brooks’ approach survived in industrial applications (iRobot, Boston Dynamics’ early work) but lost the theoretical battle.

The enactive strand of 4E has been the most theoretically developed. Francisco Varela, Evan Thompson, and Eleanor Rosch’s The Embodied Mind: Cognitive Science and Human Experience (MIT Press, 1991) is the foundational text — cognition emerges from the embodied action of an organism in its environment, structured by the organism’s history of structural couplings rather than by representations of an external world. Thompson’s Mind in Life (Belknap/Harvard, 2007) provides the rigorous philosophical extension. Hanne De Jaegher’s “Participatory Sense-Making” (Phenomenology and the Cognitive Sciences, 2007, with Ezequiel Di Paolo) develops the social-enactive case: cognition between two agents is itself a coupled dynamical process, not the meeting of two pre-formed minds. Alva Noë’s Action in Perception (MIT, 2004) and Out of Our Heads (Hill and Wang, 2009) make the case that perception is something we do rather than something that happens to us. Anthony Chemero’s Radical Embodied Cognitive Science (MIT, 2009) is the manifesto for the strongest version of the program — that representations should be eliminated from cognitive science altogether in favor of dynamical-systems coupling.

The modern AI implication is that benchmarks are constructed without affordance structure. ImageNet shows static photographs to a model with no body; the model has no relation to the depicted objects beyond recognizing them as members of a category. ARC-AGI shows abstract grid puzzles to a model with no developmental history of physical manipulation. The “general intelligence” the benchmarks claim to measure is general only across a class of de-bodied, de-coupled tasks that an enactive cognitive scientist would say are not where intelligence lives. Until the field begins evaluating cognition in coupled ecological settings — agents acting on environments they have histories with, against problems whose solutions matter to the system itself — the assessment is structurally uninformative about what an enactive theory would call cognition.

Collective and distributed cognition

Turing’s criterion is a single conversational partner. The consequence: cognition as fundamentally social, linguistic meaning as public-and-enacted, became “context” rather than constitutive. Wittgenstein’s later work — Philosophical Investigations (Blackwell, 1953) — argued that meaning is constituted in language games, public practices in which words get their meaning from their use in shared activity, not from an internal mental state in the speaker’s head. Vygotsky’s zone of proximal development pushed the same point developmentally: cognition is acquired interactionally, not internally. Edwin Hutchins’ Cognition in the Wild (MIT Press, 1995), referenced in post 1, was the canonical anthropological case study — ship navigation as a cognitive system distributed across navigator, bridge crew, chart, pelorus, and the ship itself.

Beyond Hutchins, Dan Sperber’s “epidemiology of representations” (laid out in Explaining Culture: A Naturalistic Approach, Blackwell, 1996) reframes cultural transmission as the differential survival and replication of representational variants in populations of minds — cognition at the population level, not the individual level. Steven Sloman and Philip Fernbach’s The Knowledge Illusion: Why We Never Think Alone (Riverhead, 2017) develops the experimental case that what individuals call “knowing” is mostly access to a distributed knowledge community; the felt sense of personal understanding outstrips actual personal understanding by orders of magnitude on most topics, and this is functional rather than pathological. Tomasello’s shared-intentionality work, already cited, extends the developmental-evolutionary case.

This connects directly back to Bateson’s “difference that makes a difference” definition of information that I anchored in post 1. Bateson’s information is not a payload in a head; it is a difference whose registration matters to a coupled system. In a distributed-cognition reading, the system that is doing the cognizing is the whole social apparatus, not the individual; the information lives in the differences between minds, not within each one. This is also where the I=E framework has the most natural fit, because endogenous elimination at the collective level is what shared cognitive practices are — a community continuously eliminating non-viable interpretations through interaction. The Turing filter, which evaluates one model in conversation with one human, can in principle never measure this kind of cognition.

The modern AI implication is largely unaddressed. LLMs are trained on the recorded output of collective human cognition — billions of human-hours of writing, accumulated across cultures and centuries. The model is, formally, a compression of that distributed cognitive process. Yet the model is theorized and evaluated as if it were an individual agent with individual capacities. We barely theorize multi-agent emergent cognition (post 1’s discussion of Pask’s conversation theory bears on this), even as LLMs clearly exhibit properties that only make sense as compressions of collective human cognitive output. When a model “knows” something, what is “knowing” doing in that sentence — pointing at a property of the model, or at a property of the corpus the model compressed? The Turing filter forbids the question by structurally treating the system in front of the interrogator as the thing being assessed, with no formal access to where the system came from.

What this did to the people doing the work

More uncomfortable and less discussed: the operational filter shaped the demography and the methodological repertoire of the field that downstream-inherited it. A pipeline self-selected for people comfortable with behaviorist framing — engineers optimizing measurable metrics — and selected against people trained in contemplative practice, phenomenology, comparative cognition, non-Western philosophy, or developmental biology. The selection pressure is concrete: every publication venue accepts behavioral-benchmark papers; very few accept first-person phenomenological reports as data; the introspection-based traditions have no recognized format in mainstream cognitive-science or ML conferences.

The empirical evidence for the demographic narrowing is visible in any analysis of NeurIPS, ICML, or ICLR author backgrounds across decades — the field has shifted progressively toward a narrower educational profile (CS or applied math undergrad, ML PhD, industry employment) and away from interdisciplinary entrants (philosophy, psychology, biology, anthropology). The “epistemic monoculture” concern that philosophers of science like Heather Douglas (Science, Policy, and the Value-Free Ideal, Pittsburgh, 2009) and Peter Galison (Image and Logic, Chicago, 1997) have raised about specific scientific fields applies here in a strong form: the methodological filter has produced a cognitive monoculture among researchers, which then reinforces the filter, which then further narrows the methodology.

There is a parallel worth noting in the broader replication crisis in psychology. The Open Science Collaboration’s 2015 paper in Science — reproducing 100 psychology experiments and finding that only ~36% yielded statistically significant results comparable to the originals — has been variously interpreted, and the methodological debates it provoked are not settled. What it does show, at minimum, is that experimental behavioral methodology in psychology has reliability problems of its own, which complicates any straightforward “introspection is unreliable, behavior is reliable” framing. This doesn’t vindicate introspection — that’s a separate empirical question the field largely stopped investigating before it had been settled. It does mean the historical case for displacing trained introspection wholesale rests on a comparison that has not aged as cleanly as the displacement assumed.

The consequence is that the field lost the human capacity to notice things about cognition that only trained first-person observation reveals. You cannot get that back just by funding it — it requires rebuilding trained introspective competence in researchers, which takes years per person, against an institutional structure that does not currently incentivize the investment.

What the Turing test got right

I have been one-sided so far. As with post 1, honest steel-manning is owed.

The Turing test got real things right, and they should be named. In 1950, machine cognition was not a discussable topic in serious scientific venues. The behaviorist consensus had ruled introspective and metaphysical claims about mind out of bounds; the symbolic-AI ontology that Dartmouth would name six years later did not yet exist as an institutional framework; cybernetics was the only competitor and was already getting the woolliness reputation that would eventually marginalize it. Without some operational substitute, the question “can machines think?” would have remained a barroom topic indefinitely. Turing’s test gave the field a concrete falsifier — you could imagine a machine passing or failing it, and this thought-experimental tractability was what made the field possible at all. Treating that achievement lightly because the substitute outlived its provisional status is unfair to Turing.

The test also sidestepped a metaphysical morass that would have killed serious work. The questions Turing dismissed in the opening of his paper — what does it mean for a thing made of metal and glass to “think,” what would it require for a non-biological system to have inner experience — are still open seventy-five years later, and there is no consensus on whether they are even tractable. If Turing had insisted on solving them before allowing engineering work, no modern computing-and-cognition field would exist, and we would not have the empirical basis on which we can now reopen those questions productively.

Several of the recurring critiques of the test are objections Turing already addressed pre-emptively. Searle’s Chinese Room argument (Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences, 1980) — the famous thought experiment in which a person mechanically following rules to produce Chinese responses lacks Chinese understanding, therefore syntactic manipulation is not semantic understanding — is structurally a version of Turing’s “argument from consciousness” objection, which Turing handled by arguing that demanding direct access to another’s inner experience as a criterion would force solipsism. Searle’s argument is sharper than Turing’s preemptive dismissal acknowledged, but it isn’t novel; the structural shape of the objection had been seen and addressed in 1950. Many modern critiques of LLMs — that they are stochastic parrots, that they pattern-match without understanding, that they can pass tests without thinking — are descendants of objections Turing had already considered. This doesn’t make the modern critiques wrong. It does mean that the critique-of-the-critique-of-the-test conversation has been going on for seventy-five years and has not converged, which is information about the difficulty of the question, not just the Turing tradition’s blindness.

Turing himself was explicit that the substitute was provisional. The famous prediction at the end of his paper — “I believe that in about fifty years’ time it will be possible to programme computers… to make them play the imitation game so well that an average interrogator will not have more than 70 per cent chance of making the right identification after five minutes of questioning” — has aged interestingly. By 2000 the test was nowhere close to passable in the strong form Turing meant. The current state in 2025 is harder to summarize cleanly: there is no single agreed-upon protocol for testing GPT-class systems against the original setup, and informal restatements vary widely in rigor. What can be said is that casual interrogators in unstructured chat settings frequently report difficulty distinguishing capable LLM responses from human ones, while more structured adversarial protocols continue to expose differences. The test’s metaphysical silence — what is underneath the verbal performance — remains exactly as undetermined as Turing predicted it would be, and that silence was Turing’s explicit choice. The field inherited the silence; Turing didn’t impose it.

The honest position: the test was a good move in 1950 and a constraint that should have been retired by 1970. The persistence past its natural lifespan is the field’s failure, not Turing’s. This is the same shape as the steel-manning in post 1: the cuts were not made by villains; the costs of the cuts compounded under institutional pressures Turing and McCarthy did not control.

Where does this leave us now?

There are plenty of other eliminations that happened with the choice — I have surveyed the main ones — but the synthesis is that Turing’s operationalism collapsed a multi-dimensional question. “What is a mind?” “What is its first-person character?” “How does it come into being?” “How does it relate to its body and world and others and internally?” “What other forms has it taken across species and cultures?” All of these were collapsed onto a single axis: can a candidate fool a human evaluator in text. The word Intelligence in AI is an artifact of this filter — it means whatever passes a test humans can score. Drop the filter and the term no longer has a definition; it has only a history.

The field got extremely good at the operationalized version, because of the seven decades of serious work invested since 1950, but it forgot the other dimensions existed along the way. This is not an attempt to treat Turing as a villain; he is far from that. He proposed the test as a pragmatic move around a philosophical conundrum dominating the field globally at the time, and explicitly wrote that the question “can machines think?” was too meaningless to deserve discussion. He expected the operational substitute would be temporary. It was, collectively, the field’s choice to eliminate the other dimensions, because the substitute proved convenient and expressible, and the field never replaced it. A provisional definition of mind became the permanent one. That’s what filters do when nothing replaces them.

What can be done now? For the current generation of AI systems, the filter probably cannot be undone. The architectures were designed inside it; the benchmarks were defined inside it; the institutional incentives lock it in. But the recovery of the questions — the ones the filter declared illegitimate — is happening on the margins, and the margins are widening. Consciousness research has gotten respectable again. Phenomenology is being taken more seriously by parts of cognitive neuroscience. Comparative cognition is producing results the human-shaped silhouette cannot absorb. Non-Western philosophical traditions are being recovered into the cognitive-science literature, slowly. Developmental and ecological accounts of cognition have a foothold. The next generation of machine systems — whatever they are called, assuming the names AI and AGI survive — will be designed by people some of whom have read this material and some of whom have not. Whether they recover the eliminated dimensions or build on top of the filter for another seventy-five years depends on choices being made right now, by people who may or may not know they are making them.

Naming the filter is the prerequisite for being able to choose it deliberately rather than to inherit it by default. That is what this post tries to do.