Remembering: a study in experimental and social psychology

TL;DR — In 1932 Frederic Bartlett showed, by passing a Native American folk tale ("The War of the Ghosts") through chains of Cambridge students, that human memory does not replay stored records: it regenerates content from a compressed, culturally-shaped model he called a schema, plus a few surviving fragments — and the regeneration systematically drifts toward the rememberer's expectations. The idea was unmodelable in 1932, lost fifty years to storage metaphors, and is now the most natural description we have of what a large language model does when it "recalls" something. Serial reproduction is, quietly, one of the cheapest and most rigorous probes we have for measuring the priors baked into both brains and weights — and almost nobody is using it at scale yet.

One bibliographic note before starting: the linked DOI is a 1933 British Journal of Educational Psychology item co-listed with Cyril Burt — almost certainly Burt's review of Bartlett's book, which was published by Cambridge University Press in 1932. The ideas discussed here are Bartlett's, from the book. I have not verified page-level details of the 1933 journal item.

The experiment, and what Bartlett actually claimed

The method is the children's game of telephone, done carefully. Bartlett gave a subject a short story from an unfamiliar culture, waited, asked them to reproduce it, then gave that reproduction to the next subject, and so on down a chain — serial reproduction. He also ran repeated reproduction: one person recalling the same story at increasing delays.

The results were not random decay. The distortions had structure, and always the same structure:

• Rationalization. Supernatural or culturally alien elements ("something black came out of his mouth") were dropped, softened, or reinterpreted until the story made sense to an Edwardian Englishman.
• Conventionalization. "Canoes" became "boats"; "hunting seals" became "fishing." Unfamiliar tokens snapped to the nearest familiar category.
• Sharpening and leveling. A few vivid details were preserved or exaggerated; everything else was smoothed away.
• Confabulated coherence. Subjects added connective material that was never in the original, to make the causal structure flow — and were confident it had been there.

Bartlett's interpretation was radical for its time. Memory is not a warehouse of traces. Recall is "an imaginative reconstruction, built out of the relation of our attitude towards a whole active mass of organised past reactions" — a generation process conditioned on a schema plus a handful of retained fragments. He called the drive to fit input to schema the effort after meaning. The errors are not noise; they are the fingerprint of the generative model doing the recalling.

In modern language: the brain stores a lossy, compressed representation; recall is decompression through a learned prior; and the prior's biases show up as systematic distortion. That is a startlingly complete statement of memory-as-generation, made two decades before Shannon and ninety years before anyone could build the machine it describes.

Why it could not win in 1932

Bartlett's problem was that "schema" had no formal content. He knew it himself — he complained the word was vague and used it reluctantly. There was no mathematics of compression (Shannon's 1948 paper was sixteen years away), no notion of a learned generative model, no way to say what "reconstruction from a schema" computed. Against this, the Ebbinghaus tradition — nonsense syllables, retention curves, memory as a decaying trace — was quantitative, replicable, and publishable. Rigor beat truth.

Then the computer arrived and made things worse. From the 1950s the dominant metaphor was addressable storage: encode, store, retrieve, with faithful readout as the default and distortion as an exception needing explanation. Bartlett's view — distortion as the signature of the mechanism — inverted that, and inversion is hard to sell without a working model.

And a working model was numerically out of reach. A schema, cashed out, is a generative model of everyday events and narratives — something you can only obtain by fitting an enormously flexible function to enormous data. Rough numbers: a 1932 human computer managed on the order of one operation per second; the fastest machine of the 1950s, maybe 10^4; a trained modern LLM represents roughly 10^{23}–10^{25} floating-point operations of fitting on the order of 10^{13} tokens. The gap between what Bartlett's theory required and what any machine could do was around twenty orders of magnitude. Even the concept of learning such a model — gradient descent on a distributed representation — didn't exist until connectionism in the 1980s, which is exactly when schema theory (Rumelhart, Schank & Abelson's scripts, McClelland's PDP schemata) came back from the dead.

What changed: we built the Bartlett machine

An LLM is a compressed model of text with no verbatim store. Ask it to recall a paper, a quotation, a legal case, and it regenerates the content from weights, conditioned on whatever fragments the prompt supplies. When the fragments underdetermine the answer, the prior fills the gap — fluently, confidently, and wrong in schema-consistent ways. It invents the plausible citation, the conventional plot, the API that should exist. This is not "identical" to human confabulation at the mechanism level — that would overclaim — but it is the same kind of error Bartlett catalogued: rationalization, conventionalization, confident filling-in. Hallucination is effort after meaning.

Neuroscience converged on the same picture independently. Complementary learning systems (McClelland, McNaughton & O'Reilly, 1995) split memory into a fast episodic index (hippocampus) and a slow statistical model (cortex), with replay consolidating one into the other. Recent generative-memory models make Bartlett explicit: Spens & Burgess (2024, Nature Human Behaviour) model episodic recall as reconstruction by a generative network trained on hippocampal traces, and reproduce classic distortion effects — including boundary extension and gist-based errors — as consequences of the architecture. Nagy, Török & Orbán have framed forgetting as rate-distortion-optimal semantic compression. In all of these, distortion is not a bug appended to memory; it falls out of memory being generation.

The third convergence gave Bartlett's method its missing mathematics. Griffiths and colleagues showed that serial reproduction by Bayesian agents is a Gibbs sampler: if each agent infers a hypothesis h \sim p(h \mid x_t) from the story it heard and reproduces x_{t+1} \sim p(x \mid h), the chain's stationary distribution is the prior predictive p(x). The telephone game doesn't just degrade a signal — it converges to the participants' prior. Xu & Griffiths demonstrated this empirically with humans. Serial reproduction is a prior-elicitation instrument, and it works on any system that can read and retell a story.

It is already being pointed at models: Acerbi & Stubbersfield (2023, PNAS) ran transmission-chain experiments on GPT-3.5 and found human-like content biases — preferential survival of negative, social, and threat-related information — matching the cultural-evolution literature on human chains. That is Bartlett's paradigm, run on a Bartlett machine, finding Bartlett's result.

What a serious 2026 revival looks like

The opportunity is to turn serial reproduction into a standard, quantitative interpretability protocol — a way to read out the prior in the weights and compare it, on identical footing, with the prior in a population of human heads.

The protocol: take a corpus of stories with controlled properties (schema-congruent vs. incongruent events, cultural distance from training data, causal-structure violations). For each model, iterate: present story, force genuine reconstruction — clear the context, or interpose a distractor task, or paraphrase-bottleneck through a low-capacity summary so verbatim copying is impossible — then feed the reconstruction back as the next input. Run chains of 20–50 steps, many seeds. Do exactly the same with human chains (crowdsourced; this is cheap).

What to measure:

• Attractors. Where do chains converge in embedding space? The stationary distribution is the prior; its modes are the model's schemas, made visible without opening the weights.
• Distortion vector fields. At each step, the displacement from input to output story is a sample of the prior's pull. Averaged over inputs, you get a flow field on narrative space — which incongruent elements get rationalized, in which direction, how fast.
• Convergence rate as a capacity measure. A chain that converges in 3 steps has a strong, narrow prior; 30 steps, a weak or broad one. Compare across model scale, RLHF vs. base, and against humans. The Gibbs-sampler framing makes these comparable in units of mixing time.
• Interventions. Fine-tune a schema in (a fictional culture's conventions), verify the attractor moves. This turns serial reproduction into a causal test of what a training intervention actually installed — arguably more direct than probing.

Reuse from Bartlett: the paradigm, the taxonomy of distortions (rationalization, conventionalization, sharpening/leveling — these are natural labels for automated classification), the insistence that errors are the data. Replace: his informal analysis with embedding-space geometry and the Markov-chain formalism; his n\approx tens of Cambridge students with thousands of chains across models and populations.

Status: vindicated, with the best part still open

The core claim is as vindicated as ideas get. Reconstructive memory returned via Neisser and Loftus (whose misinformation-effect work is Bartlett applied to eyewitnesses), schema theory returned via connectionism, and both the dominant artificial memory system (LLM weights) and the leading computational accounts of hippocampal function are now explicitly generative-reconstructive. Bartlett's book is among the most cited works in psychology; the modest citation count on the linked DOI reflects the 1933 review, not the book.

What is open is the methodological payoff: serial reproduction as a routine, scaled instrument for comparing priors across models, checkpoints, and human populations — and the harder scientific question of whether LLM distortion dynamics and human distortion dynamics match quantitatively (same attractors? same mixing times? same taxonomy proportions?) or merely qualitatively. Acerbi & Stubbersfield is one datapoint. Nobody has published, to my knowledge, the full flow-field version, or used chain convergence as a training diagnostic. It is cheap, it is theory-grounded, and it works identically on brains and models. That combination is rare.

Where to read it

The linked DOI (10.1111/j.2044-8279.1933.tb02913.x) is the 1933 BJEP item — read it as a period document, but the primary text is Bartlett's Remembering: A Study in Experimental and Social Psychology (Cambridge, 1932); the "War of the Ghosts" chapters remain genuinely enjoyable. Alongside it: Xu & Griffiths on serial reproduction as sampling from the prior (around 2010); Acerbi & Stubbersfield (PNAS 2023) for LLM transmission chains; Spens & Burgess (Nature Human Behaviour 2024) for the generative-hippocampus model; and McClelland, McNaughton & O'Reilly (1995) for the consolidation story that connects them. Read Bartlett last of all, and notice how little translation his prose needs.