The strongest version of the stochastic-parrot argument is easy to caricature, which is unfortunate because the original work was considerably more careful. Bender and Koller distinguished linguistic form from meaning understood through communicative intent, arguing that training on form alone does not amount to acquiring human-like grounded meaning (Bender & Koller, 2020). A year later, On the Dangers of Stochastic Parrots examined a much broader set of problems around large language models, including training data, environmental costs, bias and the risks of humans mistaking synthetic linguistic fluency for meaningful communication (Bender et al., 2021).

Those arguments remain important: a large language model has not stood in Paris, it has not experienced embarrassment, it does not acquire the word mother through anything remotely resembling the developmental history through which a child does.Nothing in the research I discuss below changes that.The slippage happens elsewhere. A warning against inferring human meaning from language behaviour gradually became a claim about the absence of meaningful internal organisation.The popular syllogism now sounds roughly like this: the model is trained to predict tokens… therefore it learns statistical correlations, therefore apparently conceptual or reasoning-like behaviour is only a more elaborate form of statistical correlation.The first proposition is true, the second is almost vacuous, the third does not follow.Calling something statistical tells us remarkably little about the organisational level at which useful explanations will eventually live. Human brains are statistical physical systems too. This observation has, so far, failed to abolish neuroscience. The mistake is subtler. We have confused the selection pressure with the solution selected under that pressure.

Interpretability research increasingly finds internal variables that generalise across contexts, participate in computation and can be manipulated to produce predictable behavioural changes: representations of board state, factual truth, abstract concepts and functional emotional states; reusable circuits; and, most recently, a restricted internal workspace associated with reportability, control and intermediate reasoning (Li et al., 2023; Marks & Tegmark, 2024; Templeton et al., 2024; Sofroniew et al., 2026; Gurnee et al., 2026).

I think the interesting implication is no longer that LLMs might secretly think like humans: it is almost the opposite. They may be the first large-scale artefacts whose functional organisation is acquired rather than explicitly authored, forcing us to discover their internal concepts experimentally after training rather than reading those concepts from their source code. That would make them something rather unusual: engineered systems with a developmental history.

And it raises a more consequential possibility.

Traditional software engineering gives us a comforting causal story: somebody writes an algorithm, the computer executes it. If the software sorts a list using mergesort, there is a perfectly ordinary sense in which the sorting procedure existed in the engineer’s design before the machine executed it.Deep learning partially breaks this relationship.Engineers specify architectures, objectives, optimisation procedures, datasets, regularisation and increasingly elaborate post-training environments. But the detailed internal representations and computational strategies that eventually satisfy those constraints are learned through optimisation rather than individually written into the system. Representation learning was developed precisely around the premise that useful latent factors can be learned from data rather than manually specified (Bengio, Courville & Vincent, 2013).This distinction deserves more attention than it gets: there is authored computation and there is acquired computation. A transformer architecture is authored. Many of the mechanisms instantiated after training are acquired.The second exists because of the first, obviously. Nothing mystical has happened. Gradient descent remains stubbornly mathematical… but causally dependent is not the same as explanatorily identical.Nobody explicitly programmed an Othello board representation into Othello-GPT, yet researchers found that information about the latent board state was encoded internally and that interventions on that representation changed subsequent predictions (Li et al., 2023).Nobody wrote a Fourier-based modular-addition algorithm into the small transformers studied in work on grokking, yet mechanistic analysis found that training had produced precisely such an algorithmic solution (Nanda et al., 2023).Induction-head research similarly identified learned circuits associated with in-context pattern completion and showed that their formation during training coincided with a sharp increase in in-context learning ability (Olsson et al., 2022).This suggests a category for which ordinary software language is becoming slightly uncomfortable.I would call these systems post-programmed artefacts. The distinction matters because source code is no longer a sufficient provenance record for behaviour. In a post-programmed system, explaining why a capability exists may require reconstructing not only the architecture that permitted it, but the developmental path through which optimisation produced it.We specify a space of possible organisations, optimisation selects and constructs one trajectory through that space, and then, in a slightly humiliating reversal of the normal relationship between programmers and software, we open the trained system and try to work out what happened.

This is where I think both sides of the current debate use an embarrassingly weak ontology.One side finds a model producing behaviour associated with reasoning, emotion or planning and immediately reaches for human psychological vocabulary. The other notices that the implementation consists of vectors and matrix operations and concludes that psychological vocabulary must therefore be metaphorical.Neither criterion is particularly scientific.A better question is whether a higher-level variable does explanatory work.Daniel Dennett’s notion of real patterns was built around a related intuition: higher-level entities need not disappear merely because a lower-level physical description exists, provided the higher-level pattern captures genuine regularity and improves explanation or prediction (Dennett, 1991). Recent philosophy of science has explicitly revisited this idea in the context of artificial intelligence and large neural networks (Millhouse, Petersen & Ross, 2026).I would push the criterion further for learned systems. Call it the mesoscopic cognition hypothesis.I would reserve the word cognitive for a stronger case. A mesoscopic variable should not merely correlate with behaviour or survive intervention: it should be reused across contexts, support counterfactual control, participate compositionally in several computations, and transfer across tasks that depend on the same underlying abstraction.The question is not simply whether the system represents something: it is whether that representation becomes a reusable piece of its internal problem-solving economy.The last condition matters most: correlation gives us a label, intervention starts giving us a mechanism.This is why some recent results are qualitatively more interesting than another probe discovering that a hidden layer contains information about nouns.Marks and Tegmark found a direction associated with the truth or falsity of simple factual propositions that transferred between datasets; interventions along the direction could push the models studied towards treating false statements as true and vice versa (Marks & Tegmark, 2024).Work on Claude 3 Sonnet extracted highly abstract features that could respond to the same conceptual content across different contexts and, in some cases, languages or modalities; artificially activating particular features could alter downstream model behaviour (Templeton et al., 2024).In 2026, researchers at Anthropic reported internal representations corresponding to emotion concepts in Claude Sonnet 4.5 that generalised across contexts and causally influenced preferences and behavioural tendencies. They deliberately use the term functional emotions, while explicitly denying that the evidence establishes subjective emotional experience (Sofroniew et al., 2026).Another 2026 study identified a restricted set of representations, called J-space, associated with verbal report, deliberate control, flexible downstream use and intermediate steps in silent reasoning, drawing a functional comparison with global-workspace theories while again stopping short of claims about phenomenal consciousness (Gurnee et al., 2026).These results are young: some come from the same research group.Methods such as sparse autoencoders and circuit tracing recover only partial approximations of the underlying computation. Replications will undoubtedly make several current interpretations look charmingly primitive.That does not yet mean we have discovered the model’s natural ontology. The features recovered by our interpretability tools may partly reflect the resolution at which we choose to inspect the system.Not a neuron, not an output: something in between. Perhaps artificial cognition has no privileged conceptual granularity. What appears to us as a concept may instead be one usable resolution of a richer representational geometry.That middle scale is where biology has organs, economics has firms and thermodynamics has temperature.Perhaps artificial cognition has its own legitimate macroscopic objects too.

Once internal structure can be acquired, another assumption begins to fail.A model is not only characterised by what it contains at the end of training: it has a history of becoming. Studies of induction heads found that particular computational structures form during identifiable periods of training alongside changes in in-context learning behaviour (Olsson et al., 2022).This idea has developed into a nascent field sometimes called developmental interpretability, which studies how internal structures and computational regimes arise across training rather than treating a completed neural network as a frozen object to be reverse-engineered after the fact. The explicit analogy is developmental biology: understanding the mature system partly through the process by which its organisation formed.The analogy became considerably less poetic in 2026. Chen, Luo and Pan introduced mechanistic data attribution, tracing interpretable components back to influential training examples. In Pythia models, altering small subsets of high-influence training samples could modify the emergence of particular heads, while interventions targeting induction-head development produced corresponding changes in in-context learning capability (Chen, Luo & Pan, 2026).Notice how strange the vocabulary is becoming for something we continue to call software. Formation. Emergence. Development. Intervention during development. Mature internal structure.Anthropic’s interpretability team did not accidentally title its 2025 circuit-tracing work On the Biology of a Large Language Model. The project examines learned internal mechanisms by tracing functional pathways through the network rather than deriving them directly from a programmer-written algorithm (Lindsey et al., 2025).I would not take the biological analogy literally. An LLM does not metabolise, reproduce or heal. Training is not embryogenesis.But the methodological shift is real: for ordinary software, asking how did this algorithm develop? would be a peculiar question. For a learned model, it may become indispensable.That gives us a new distinction: architecture tells us what could develop; training history helps explain what did. And that makes the final model not simply a program, it is also the residue of a developmental trajectory.

This brings me to the claim I find much more interesting than whether GPT-7 deserves the verb think.Cognitive science has obviously never had literally one specimen. It studies children and adults, humans and non-human animals, healthy and damaged brains, different species and radically different behavioural ecologies.An octopus is not simply a damp human with more arms. But every naturally occurring cognitive system we have studied is the product of biological evolution. It is constrained by metabolism, survival, reproduction, sensorimotor embodiment and some variation on living tissue.Artificial neural systems arrive through an entirely different causal route.Their architectures are engineered.Their representations are acquired through optimisation.Their experience is mostly data.Their objectives are imposed from outside.Their substrate is artificial.Their developmental timescale is absurd.They are not an independent replication of human cognition.And that is precisely what makes them useful.The novelty is not simply that artificial systems provide another kind of cognitive architecture (comparative cognition already gave us other architectures): the novelty is that we can rerun their development. We can change the training distribution. Remove an experience. Alter an objective. Save developmental checkpoints. Retrain from a different initialisation. Intervene on a representation halfway through learning and observe what replaces it.We cannot rerun the evolution of the hippocampus ten thousand times under controlled conditions. We can, at least in principle, do something disturbingly close to that with artificial systems.There are already tantalising, if disputed, hints.The Platonic Representation Hypothesis proposed that increasingly capable models might converge towards shared representational structure across architectures and modalities (Huh et al., 2024).Later work has significantly qualified the stronger version of that claim: once scale-related confounds are controlled, global geometric convergence appears much weaker, while agreement in local relational structure survives more convincingly (Gröger, Wen & Brbić, 2026).Even more recent work suggests the distinction may be between convergence of relations and convergence of exact metric geometry: independently trained models can preserve similar structural neighbourhoods without encoding those relationships in precisely the same coordinates (You et al., 2026).I think this is far more interesting than asking whether two models contain the same vector for cat.If independently optimised systems repeatedly recover similar relations, the invariant may live one level above implementation. And if human cognition recovers some of the same relational structures, the question becomes genuinely strange. Perhaps a concept is not fundamentally a human psychological object at all. Perhaps it is a recurrent solution class: a structure that sufficiently capable learning systems rediscover when faced with the same informational constraint.If so, the interesting invariant would not be the representation itself, but the problem that keeps making different systems invent functional equivalents of it.That possibility changes the role of artificial intelligence in cognitive science: AI stops being merely something cognition tries to explain, it becomes a second place in which candidate laws of cognition can be tested.

This is where language becomes dangerous.Say that a model has an internal representation and nobody panics. Say it has a belief and half the room imagines consciousness while the other half begins composing a rebuttal about matrix multiplication. The disagreement often arrives before the definition.There is already a research programme explicitly described as machine psychology, using methods and concepts from cognitive science to characterise stable behavioural and computational properties of artificial systems while warning against assuming that human and artificial mechanisms are identical (Hagendorff et al., 2023).I think the more radical version should go in the opposite direction. We should not begin by exporting human psychological categories into machines. We should begin by discovering machine-native functional categories, then ask whether any of them map onto categories used in biological cognition. The comparison should therefore run in both directions. We should not merely ask whether machines possess human categories such as working memory or belief. We should ask whether machine-native categories reveal distinctions that human psychology has never needed to name.That distinction matters.An internal variable need not correspond neatly to human working memory for us to discover that it performs temporary context integration. A model need not experience an emotion for an emotion-like representational state to causally reorganise its behaviour. A model need not possess human propositional belief for some internal state to track factual truth and influence subsequent processing.The scientific question is not whether the vocabulary sounds anthropomorphic.It is whether the proposed variable survives experimental attack.Can we identify it independently?Can we predict when it appears?Can we suppress it?Can we amplify it?Does doing so alter the behaviour the theory says it should alter?Does the same functional structure recur elsewhere?If the answers become yes, continuing to insist that only the microscopic description is “real” stops being scepticism, it becomes reductionism with excellent branding.This is what I mean by mesoscopic psychology.Not psychoanalysis for GPUs: a science of the causal structures between artificial neurons and whole-system behaviour.

There is an organisational consequence to all of this which I think is still badly underestimated.Software engineering historically derives much of its authority from legibility. The artefact may be complicated, but in principle the code specifies the procedure.Machine learning altered that relationship, and foundation models push it much further.The people who build a model can know the architecture, loss function, optimiser, training data pipeline and post-training regime while still being unable to give a complete mechanistic explanation of how a particular capability is internally implemented.That is why interpretability researchers probe, ablate, stimulate, trace and compare. Those verbs belong as comfortably to experimental neuroscience as to traditional software engineering.And this creates what I would call an epistemic inversion of engineering.Classical engineering proceeds from mechanism to artefact. Learned-system engineering increasingly proceeds from conditions to artefact to mechanism. That is not merely opacity. It is a reversal in the epistemic order of engineering.Moreover, this distinction changes safety: if important internal structures emerge through developmental trajectories, endpoint behavioural evaluation alone can never tell us everything we might want to know about how the behaviour is produced. Research tracing circuit formation and training-data influence suggests that properties of the mature network can depend on identifiable developmental processes during training (Olsson et al., 2022; Chen, Luo & Pan, 2026).A mature discipline may therefore need something closer to a developmental record of a model, not merely a final benchmark sheet.Which representations appeared?When did particular circuits become functional?Which training regimes promoted them?Which survived post-training?Which apparently disappeared from behaviour while remaining internally available?This is still a research proposal, not current engineering practice; but if the developmental view is right, the provenance of a model eventually needs to describe more than its data, architecture and evaluations. It may eventually need something closer to a developmental provenance: a record not only of what the system was built from, but of how its functional organisation came to exist. For learned systems, provenance may need to become causal rather than merely documentary.

This, to me, is the part worth taking seriously. For seventy years, AI research has frequently judged progress by proximity to us.Can it converse like us?Reason like us?See like us?Plan like us?Pass our examinations?Beat our experts?The comparison made sense because humans were the only obvious existence proof for broadly capable intelligence.But that framing may now be becoming scientifically constraining.If artificial neural systems develop representations, circuits, intermediate functional states and developmental trajectories that are genuinely explanatory yet structurally unlike ours, the most interesting question will no longer be how close they are to human cognition.It will be what both systems had to invent, despite getting there differently.And that suggests a different question for cognitive science:Which structures belong to a particular mind, and which belong to the problems minds must solve?Is working memory a biological workaround, or a recurrent solution to limited computation?Is attention a metaphor borrowed from psychology, or a functional necessity for selective information processing?Are concepts human categories, or attractors produced by compression under structured prediction?Is metacognition inseparable from conscious self-reflection, or might some functional analogue emerge whenever a system must model the reliability of its own computations?These are different questions from asking whether an artificial system possesses our version of memory, attention or concepts. They ask whether sufficiently different learning systems, exposed to comparable computational constraints, independently rediscover functionally equivalent solutions.We do not yet know.That is precisely why the comparison matters.Artificial intelligence could become to cognitive science what synthetic chemistry became to chemistry more broadly: not merely the manufacture of useful artefacts, but a way of testing which structures are possible and which principles constrain them.Except the artefacts answer back, which was probably unnecessary.I suspect the stochastic parrot debate is approaching the end of its useful life, not because Bender and colleagues were naïve: quite the opposite.The original warning was epistemically sophisticated: linguistic behaviour does not license unlimited claims about meaning, understanding or mind (Bender & Koller, 2020; Bender et al., 2021).We should keep that discipline, but we should apply it symmetrically.If fluent language is insufficient evidence for human-like cognition, matrix multiplication is equally insufficient evidence for the absence of higher-level cognitive organisation.Both are shortcuts from one observational level to another.The new evidence suggests a more interesting middle: learned systems can develop internal representations that abstract beyond individual training examples.Some of those representations generalise.Some organise computation.Some can be manipulated causally.Some emerge at identifiable stages of training.Some may recur, at least relationally, across independently trained systems.And none of that requires them to become miniature humans.I think this leads to a proposition considerably more interesting than LLMs understand language: artificial intelligence may let us test which structures of cognition are substrate-specific and which are problem-induced.If that proposition is correct, then we need to stop treating psychology as the study of one implementation and computer science as the study of another.The interesting science will live across the boundary: it will ask which functional structures are specific to brains, which are specific to artificial networks, and which are attractors that sufficiently capable learning systems keep rediscovering because the problems they face demand them. And unlike almost every previous object of cognitive science, these systems allow us to manipulate the developmental conditions that produce those structures.Find a proposed machine-native cognitive variable.Show that it generalises beyond the examples used to discover it.Show that it performs causal work.Trace how it develops.Look for functional equivalents in independently trained systems.Then look for analogues in biological cognition.If the structure disappears under those tests, discard it.If it survives, promote it into the ontology.This does not mean the LLM has a mind, but the mechanism earned the description.We may eventually discover that many apparently psychological properties of LLMs collapse under scrutiny.Good.But we may also discover that cognition is a much larger design space than the single strange biological solution through which we first encountered it.The embarrassing possibility is not that the stochastic parrot turned out to have a psychology.It is that, by building minds we can rerun, we may finally be able to discover which parts of psychology were never specifically ours.