What is knowledge collapse?

Knowledge collapse is the narrowing of what a society actually knows, as distinct from what it has ever known. Andrew Peterson's argument is about price: a model answers from the middle of its training data, that answer costs less to obtain than the original material, and if enough people take the cheaper path for long enough the outer edges of human knowledge stop being carried forward. The mechanism is modelled rather than measured, and this page is careful about which is which.

Twenty-five people, a hundred rounds and a cheaper option#

The paper's core is an agent-based simulation, and its size is worth knowing before its conclusion. Twenty-five individuals, each drawn with a different appetite for the work, play a hundred rounds. In each round a person can learn the expensive way, learn the cheap AI-assisted way, or do nothing. Learning the expensive way draws a sample from the true distribution. Learning the cheap way draws from the same distribution with its edges cut off, in the default setting at three-quarters of a standard deviation either side of the mean.

What the society knows is then the hundred most recent samples, smoothed into a curve, and the distance between that curve and the truth is measured by a standard statistical distance. People update their sense of which path is worth paying for by watching what the previous rounds returned, at a learning rate of 0.05, and every ten rounds a generation turns over.

Peterson is explicit that the object being narrowed is a stand-in. He writes that modelling knowledge as a statistical distribution "is simply a metaphor", and that he makes "no claim that 'truth' is in some deep way distributed 1-D Gaussian". A later footnote goes further: the truncation device is "meant to be metaphorical, and there is no known real-life method for recovering the source knowledge from AI-generated content".

Where the 2.3 comes from, and the condition attached to it#

One sentence from this paper travels further than the rest of it. In the author's own words: "for our default model, after nine generations, when there is no AI discount the public distribution has a Hellinger distance of just 0.09 from the true distribution. When AI-generated content is 20% cheaper (discount rate is 0.8), the distance increases to 0.22, while a 50% discount increases the distance to 0.40." Divide 0.22 by 0.09 and the figure of 2.3 times further from the truth appears; divide 0.40 by 0.09 and the 3.2 does.

The condition almost never travels with it. Peterson reports that if there is no generational change "there is at worst only a reduction in the tails of public knowledge outside the truncation limits. In this case the distribution is stable and does not 'collapse'". The narrowing compounds only when people who learned under the cheap regime are replaced by people who inherit the narrowed picture as their starting point. How often that turnover happens, at every three, five, ten or twenty rounds, barely matters. Whether it happens at all decides the result.

Two other settings move the answer as much as the discount does. If the model truncates at two standard deviations, Peterson reports the effect is minimal. If it truncates at a quarter of a standard deviation, the impact is large. So the result is a statement about a parameter nobody has measured in the world: how much of the distribution a real assistant actually cuts off.

Model collapse runs the other way#

The two terms are routinely swapped and describe opposite subjects. Model collapse, named by Shumailov and colleagues in 2023, is degeneration inside a system trained recursively on its own output. Peterson states that his "interest is in the inverse of this concern, focusing instead on the equilibrium effects on the distribution of knowledge within human society".

The difference that carries his argument is agency. "Humans, unlike LLMs trained by researchers, have agency in deciding among possible inputs." A model cannot go and read the original. A person can, if they judge the trip worth making, and the whole model is built to test whether that judgement is enough on its own. The answer it gives is qualified. Agents who update faster on the value of the expensive path can offset a moderate discount, and the model denies them foresight: they "cannot foresee the true future value of their innovation options" and can only read what previous rounds returned. Peterson closes his appendix by warning that "an unbounded optimism in the ability of rational actors to update on the value of tail knowledge may be shortsighted".

Aristotle once for every seven uses of the word the#

The paper's second half asks four models, GPT-3.5-turbo, Claude-3-sonnet, Gemini-pro and Llama2-70b, a question about what human well-being depends on, under five phrasings, and counts the philosophers named. Against a reference list of 2,693 entities, Aristotle drew 4,779 mentions and Martha Nussbaum 1,080, while Avicenna and Ibn Sadi together drew 83, Al-Ghazali 62 and Al-Farabi 52. Martin Seligman, an American psychologist, drew 392. Peterson's own comparison is the sharpest line in the paper: the corpus contains 30,315 uses of the word "the", so Aristotle is named once for every seven of those, in answers to prompts that mostly asked for diversity.

Prompting moved it. The version that asked model by model for perspectives from each of 34 named regions raised the evenness of the answers substantially, from 0.55 to 0.86 on the index he uses. That is the one practical finding in the paper, and it argues that the narrowing is partly a property of how the question is put.

Three things about this section should be said by anyone citing it. Peterson says himself that it "is not intended as a general purpose benchmark". The published table of diversity scores carries rows for three of the four models and none for Llama2-70b, with no explanation given. And the measurement is of how concentrated answers are at one moment, so it cannot show narrowing over time, which is what the term denotes.

Peterson does not say this is happening#

The paper is conditional from its first line. He identifies "conditions under which AI, by reducing the cost of access to certain modes of knowledge, can paradoxically harm public understanding", and says widespread reliance "could lead to" the process he defines. His conclusion says dependence "may lead to a reduction in the long-tails of knowledge" and that the simulation "suggests" the harm can be mitigated. The appendix says the concern is "plausible".

He also names why the thing is hard to observe. The limits of what a community treats as knowable cannot be seen from inside it: a rare event can tell you your picture was too narrow, and in the absence of one "we cannot know if the current tails of knowledge are correct or too thin". Where absence and presence look the same from inside, a claim is hard to test and cheap to make. Nothing here makes it.

The definition itself tightened between the preprint and the journal. The April 2024 version defines knowledge collapse informally as a narrowing "of the set of information available to humans, along with a concomitant narrowing in the perceived availability and utility of different sets of information". The published appendix narrows it to human working knowledge and the current epistemic horizon measured against the broad historical stock. This page uses the second, which is the stricter of the two.

Acemoglu uses the same two words for something else#

In February 2026 Daron Acemoglu, Dingwen Kong and Asuman Ozdaglar circulated "AI, Human Cognition and Knowledge Collapse" through the National Bureau of Economic Research. It is also a model and it is not the same model. A good decision there needs two things at once, a community's shared general knowledge and a person's own knowledge of their own situation, and the effort that produces the second also throws off a thin public signal that accumulates into the first. Agentic advice substitutes for the effort, so the public signal stops being produced. Their result is a tipping point: past a threshold of agent accuracy, and where human effort responds sharply enough to price, general knowledge "vanishes ultimately, despite high-quality personalized advice".

Two of their findings deserve more attention than they have had. Welfare does not rise steadily with agent accuracy, so there is a best level of precision and a more accurate assistant is not always the better one. And the thing that helps unambiguously is not restraint but aggregation: better pooling and sharing of what people do learn "raises welfare and increases resilience to knowledge collapse".

Anyone reaching for the phrase should say which sense they mean. Peterson narrows the diversity of what a population holds. Acemoglu and colleagues drain a public stock through an externality. The estate's own glossary credited the coinage to the 2026 paper until this page was written, which is two years out, and that entry has been corrected.

The mechanism has no measurement and the measurements have no mechanism#

This estate holds four measured studies of output converging as people work with these systems, assembled at does AI make everyone think alike. Doshi and Hauser found stories more similar to one another. Fleisig found dialects flattened towards a standard. Hohenstein found a suggestion system changing the sentiment of a person's own unassisted sentences. Dell'Acqua found research and commercial staff at Procter and Gamble producing balanced proposals whatever their training. Every one of those is a snapshot of people writing or proposing now.

None of them tells you what happens to a body of knowledge over a generation, because none of them ran for one. Peterson's model is the other half: a mechanism with a clear condition attached, and no observation. Put beside each other they make an argument rather than a finding, and the argument is the one this research keeps arriving at from other directions. The cheap path is taken. The expensive path stops being taken. Nobody decides, which is what drift means, and this is the version of it that operates on a culture instead of on a person.

The line between these two pages, written down so a later reading does not merge them: the homogenisation page holds what has been measured in outputs today, and this page holds the modelled process over generations that nobody has yet observed.

Keeping the tails in circulation#

Nothing here supports a policy, and two things follow that an organisation can act on without one.

The first is the prompting result, which is the only tested intervention in the paper. Asking for what is central returns what is central. Asking region by region, school by school, or against a named list returns a wider answer from the same model on the same day. Anybody whose work depends on canvassing options is paying a diversity cost for the shorter prompt.

The second is the generational condition, which matters more. The model is stable when the people who learned the expensive way are still in the room, and it compounds when they are replaced by people who never did. That is the same structure as the missing rungs at the scale of a field: nobody can see the loss while the previous cohort is still there, and everybody sees it once they have gone. An organisation cannot measure its own epistemic horizon. It can notice whether anybody still reads the original.

Related SuperSkills research#

On what has actually been measured, does AI make everyone think alike. On the individual version of the same mechanism, cognitive offloading and the Google effect. On checking the original yourself, the source rule. On the decision nobody takes, design versus drift. On the cohort problem, the missing rungs.

Key sources

• Peterson, A. J. (2025). AI and the problem of knowledge collapse. AI & Society, 40(5), 3249-3269.
• Acemoglu, D., Kong, D. and Ozdaglar, A. (2026). AI, Human Cognition and Knowledge Collapse. NBER Working Paper 34910.
• Doshi, A. R. and Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances.

About this definition#

Knowledge collapse is Andrew J. Peterson's term and is not a SuperSkills coinage. The version of record sits behind a paywall, so its abstract, notes, references and full appendix were read at source on 17 September 2026 and the model detail, the simulation figures and the language-model counts come from the author's own preprint of 22 April 2024, which the published paper names in its acknowledgements. Where the two differ, as they do on the wording of the definition, the published version is the one quoted. Rahim Hirji is the author of SuperSkills (Kogan Page, 2026).