Alison Gopnik on Childhood Learning, AI as a Cultural Technology, and Rethinking Nature vs. Nurture

Alison Gopnik — developmental psychologist at the University of California, Berkeley, author of The Philosophical Baby and The Gardener and the Carpenter — joins Tyler Cowen in Ep. 265 to argue that young children learn the way scientists do, that large language models are a ‘cultural technology’ rather than a new kind of mind, and that the whole nature-versus-nurture framing is the wrong question to be asking.

Key ideas

1. Children learn like scientists, building causal models from sparse data. Gopnik’s central claim is that babies and toddlers solve the same problem a scientist faces — getting from ‘a bunch of photons at the back of our retina’ to knowledge of how the world works — and they do it the same way: by inferring what hidden structure out there could have produced the pattern of evidence they see. Causal model means a picture of what causes what (fire causes smoke, not the reverse); children form and revise these from very little data, the way good computational models of scientific theory-change now describe.
2. Both children and scientists are Bayesian — in practice, not in talk. Bayesian learning means updating a belief in proportion to the strength of new evidence, weighed against how confident you already were (your prior). Ask a three-year-old to explain conditional dependencies and you get nonsense; watch what they actually do and they update rationally. Children are often the better Bayesians, because a scientist with a strongly held prior — a lot of past confirmation — is rational to resist a little contrary evidence, whereas a child with a flatter prior follows the data.
3. AI is a cultural technology, not an autonomous mind. Generative AI, Gopnik argues, belongs with print, writing, libraries and internet search: it is a new way of getting information from other people, the human ‘superpower’. An LLM trained on what intelligent humans have written gives you ‘a summary of what all the people on the net have said’. A genuinely new intelligent agent is a kitten; it will not change the world. A new cultural technology — the printing press — will. The ‘golem’ narrative of a non-living thing given a mind, for good or ill, is the wrong picture.
4. Exploration versus exploitation explains why young children roam. Borrowed from computer science, exploration means trying wildly varied options to discover what is out there; exploitation means refining and using what you already know. Gopnik likens early childhood to a ‘high-temperature search’ (the simulated-annealing idea: start with wild, random, out-of-the-box moves, then cool off and fill in detail). A four-year-old eating an avocado with a spoon mostly experiments — banging, flipping — rather than getting food in. Free of grant proposals, children can stay in the wild space all the time.
5. What today’s AI still cannot do, a two-year-old can. The gap, Gopnik insists, is intervention in the real world: designing an experiment, acting on the world, getting fresh data, and discovering something genuinely new that nobody around already knew. Robotics — going out and manipulating physical reality — is exactly where LLMs fall down and toddlers excel. She calls text-only AI ‘Derrida’s revenge’: endlessly generating more text from text, without ever making contact with an external reality.

Content

Children as scientists, scientists as children

Gopnik’s research began with a puzzle that ran in both directions. Saying children learn like scientists is only useful if we know how scientists learn — and when she started, the philosophy of science held, after Kuhn, that there was nothing systematic to say, only sociology. That changed. Computational models of theory-change emerged that turn out to fit children too. The shared core is causal inference: looking at data and working out what structure in the world could have produced this pattern. Scientists know about quarks; children know about people and things; both extract a ‘world model’, as the AI people now say, from photons and air-pressure disturbances. Crucially, scientists run on Pleistocene brains — whatever lets a child do this is what a scientist also draws on, beneath the institutions and grant cycles.

Bayesian in practice, and why kids are better at it

Cowen presses the obvious objection: scientists do not look Bayesian — they are stubborn, and when they revise they move in a predictable direction, which a true Bayesian random walk should not. Gopnik agrees the surface looks unrational but locates the rationality in practice rather than introspection. A peaked prior — a belief with much past confirmation — should resist a little new evidence; that is correct updating, not stubbornness. Children, with flatter priors, weight evidence more and so handle unusual outcomes better. She adds a social twist: a field populated by both prior-followers and evidence-followers can collectively reach the right answer, with no fixed rule for when any individual should abandon a theory.

Annealing, the avocado, and the value of fishing expeditions

To explain the difference between a child’s search and a scientist’s, Gopnik reaches for simulated annealing. Low-temperature search nudges what you already believe a little and checks whether it fits the data better — Cowen’s example of an economist edging, decade by decade, toward thinking the money supply matters. High-temperature search bounces randomly around the space of possibilities, trying wild things. The annealing recipe is to start hot and cool down. A four-year-old is unmistakably the noisy, bouncy, random creature. Her teaching video shows her son confronted with a whole avocado and a spoon: instead of eating, he bangs it, turns it over, runs every operation a spoon-and-avocado afford. Science under-credits this. The ‘fishing expedition’ we dismiss is, she suspects, how much real discovery happens — scientists experiment first, then write the grant to fund what they have already informally done.

The Friston objection and the role of intervention

Cowen raises Karl Friston’s free-energy account — minds as surprise-minimisers, interpreting data and acting to reduce prediction error. Gopnik finds it underspecified: intuitively pointed in the right direction, heavy with mathematics, but hard to tie to an empirical programme, and at risk of accounting for everything. She accepts that learners attend to violated predictions. But she pushes past surprise to intervention: you see something surprising, then you do something in the world — reproduce it, vary it, find its parameters — and that action is what drives theory-change. Given $100 million, she would not buy expensive kit (developmental psychology runs on little chairs and a $20 toy) but would put GoPros on many babies, à la Mike Frank, and mine the data for the hidden systematicity in how an infant’s actions shape the evidence it then receives.

Consciousness, memory, and the spotlight that narrows with age

Gopnik treats consciousness the way nineteenth-century biology eventually treated ‘life’: not one thing but many processes, so ‘what is consciousness?’ is likely the wrong question. The professorial prototype — the focused introspection of someone solving a problem at a desk — is, she argues, almost the opposite of awareness, because that focus blocks out everything else. Babies are conscious of everything around them; their brains are plastic, flooded with novelty, attending broadly rather than narrowly. The adult traveller in Paris, vividly awash in a new place, is the closest grown-up analogue. Weak episodic memory or aphantasia does not reduce consciousness; if anything, less autobiographical compression keeps a young child more present. Aphantasia (her friend Ed Catmull, the Pixar co-founder, has it) shows that the mental image you ‘see’ while solving a problem is epiphenomenal — a side effect of activating visual cortex, not the machinery doing the work.

AI as cultural technology — and the long argument with Cowen

The episode’s spine is a sustained, friendly disagreement. Gopnik, drawing on her Science paper with Henry Farrell and James Evans, frames generative AI as a cultural technology: a means of transmitting what other humans already know, kin to print, libraries and search, not a new agent. Cowen presses hard — reasoning models prove theorems, make small scientific discoveries absent from the internet, beat good lawyers and economists on novel problems. Gopnik holds her line: reasoning models still pick out and reproduce patterns of text, including patterns of reasoning; impressive, but not what humans or even animals do. She would be persuaded by an AI that designed an experiment and learned something new about the world that nobody around it knew. Her library analogy reframes the lawyer case: the legal code already holds all the lawyers’ prior work — accessing it faster makes law easier, but you would not call the legal code a lawyer. The human capacities are the ones that go beyond extracting information from other people.

Nature, nurture, and the carpenter versus the gardener

Gopnik thinks nature-versus-nurture, like ‘life’ and ‘general intelligence’, is a lay frame that dissolves on inspection. A single gene can make a disorder ‘100 percent nature’ (it depends on that gene) and ‘100 percent nurture’ (remove the dietary trigger and it vanishes) at once. Drawing on Eric Turkheimer, she notes twin correlations rise with wealth and fall with poverty, because in a poor family small environmental differences swing the outcome more. Her own contribution concerns caregiving: a protective caregiver increases variability. As in a sheltered garden, more plants of more kinds can thrive — which is why caring families may make siblings more different, not more alike, and why nurture’s effect on the standard deviation rather than the mean is invisible to a standard twin study. This is the thesis of The Gardener and the Carpenter: parenting is gardening, cultivating a rich and protected space for variety to emerge, not carpentry, building a child to a predetermined specification. She reads her own family — six siblings in eleven years, including Warhol biographer Blake and the New Yorker’s Adam, all ‘foxes’ — as caregiving producing divergence, not convergence.

Schooling, Goodhart’s law, and apprenticeship

For children under about seven, Gopnik endorses warm, play-based, inquiry-led learning. For school-age children, the task shifts from exploration to acquiring skills for exploitation, and the right model is intuitive apprenticeship: do the thing, get feedback from a teacher who shows you examples (and is sometimes bluntly critical), do it again — the way we teach music and sport, but almost nothing else. Imagine teaching baseball as we teach science: lectures on great games, no actual play until graduate school. The current system is a textbook Goodhart’s law failure — optimise the proxy (school performance) and children become superb at school while losing the underlying capacity it was meant to track. First-year students, masters of test-taking, freeze when asked to design an experiment of their own.

Related

• Alison Gopnik — speaker; developmental psychologist at UC Berkeley
• Tyler Cowen — host
• Paul Bloom — speaker; fellow developmental psychologist on Against Empathy and infant cognition
• Paul Bloom on the Psychology of Children, and the Morality of Empathy and Disgust — overlapping Conversations with Tyler episode on child cognition and LLMs
• Large Language Models — concept; Gopnik’s ‘cultural technology’ framing of LLMs
• Consciousness — concept; her claim that babies are more conscious, not less
• World Models — concept; the structured model of the world both children and scientists build from data
• Jagged Intelligence — concept; the uneven competence of AI that reasons yet fumbles basic arithmetic
• Hallucination — concept; why reasoning models still fabricate