DataPacts

Model Collapse and the Value of Human Data

Checked

At a glance

Model collapse is the loss of quality and diversity that research reports when models are trained, round after round, on data that earlier models generated, and it is one argument for the continued value of human-made data.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

Shumailov and colleagues described the effect in Nature, published online on 2024-07-24, and two preprints of 2024 disagree on how much real data prevents it.

This page sets out the record within AI training data licensing: the three papers with their venue status, the counter-evidence, what the record does not show, and what buyers' own listings say about generated records.

No measurement of collapse in a deployed frontier model is on record as of 2026-11-16, and no price premium for human-made data is recorded.

What model collapse means: models trained on their predecessors' output

Model collapse describes what happens when generated data, not human-made data, becomes the training input of the next model generation. The research reports a loss of quality and diversity in the models that follow, and the term covers a training-data effect, not a software fault.

The loop has three steps in plain words. A model is trained, the model generates data, and that generated data trains the next model, which then generates data for the one after it.

Shumailov and colleagues published the result in Nature, volume 631, pages 755 to 759, online on 2024-07-24, after a preprint of 2023-05-27 (arXiv 2305.17493).

Collapse is a property of training data over generations, which is why the research frames it as a question about data and not about any one model. It is not the same as a model drifting after deployment, which is a monitoring question this page does not cover.

The word "human-made" in this page means records produced by people in the course of work, such as charts, tickets, code, recordings and judgments, as opposed to text, images or code that a model produced.

The question matters on a reference about data acquired for AI training because developers choose between generating data and acquiring it. The parent page covers synthetic data for AI training and its limits, and this page covers the part of that question that concerns what happens when generated data feeds back into training.

The research record on model collapse, paper by paper

Three papers anchor the record as of 2026-11-16: one peer-reviewed in Nature and two arXiv preprints that reach different results. The table lists them by date.

Date Paper Venue status Setting Finding in the record Source
2024-04-01 Gerstgrasser et al., arXiv 2404.01413 preprint (no venue in the record) real and synthetic data accumulate across iterations test error stays under a finite upper bound independent of iterations arxiv.org
2024-07-24 Shumailov et al., Nature 631, 755 to 759 (preprint arXiv 2305.17493, 2023-05-27) peer-reviewed journal models trained recursively on generated data names and describes model collapse nature.com
2024-10-07 Dohmatob et al., "Strong model collapse", arXiv 2410.04840 preprint (no venue in the record) regression as little as 1 % synthetic data can trigger collapse arxiv.org

Three papers in the record as of 2026-11-16; the two preprints reach different results under different setups.

Venue status tells a reader how far a result has been checked. A peer-reviewed journal paper has passed the journal's review, while a preprint is a version posted before or without review, and neither of the two preprints carries a journal or conference reference in the record.

Other collapse papers exist, but none is in the record, so none is named here.

The table reports what each paper states and does not rank them. A reader should also note the order of the dates: the accumulation preprint appeared first, the Nature paper in the middle, and the one-percent preprint last, so the later paper is not a reply to the earlier one in any sense the record states.

Replacing or accumulating: the counter-evidence on mixing generated and real data

The research divides on one question: whether keeping real data in the training mix prevents collapse. The two preprints answer it differently, and both are in the table above. The distinction is between replacing real data with generated data at each round and keeping real data while adding generated data to it.

Dohmatob and colleagues (arXiv 2410.04840, 2024-10-07) report that as little as 1 % synthetic data in the training set can trigger collapse in a regression setting. The authors study a mathematical setting, and a model trained on a small share of generated data is enough to show the effect in that setting.

The record does not state that the same share has the same effect in a large language model, and the page does not extend the result beyond the setting.

Gerstgrasser and colleagues (arXiv 2404.01413, 2024-04-01) report the opposite result for a different setup. When real and synthetic data accumulate across iterations rather than replace each other, test error stays under a finite upper bound that does not depend on the number of iterations.

The setups differ, regression in one case and accumulation across iterations in the other, and neither paper is peer-reviewed in the record. The page shows the disagreement and does not resolve it: it supports neither "accumulation solves collapse" nor "1 % always collapses".

What the collapse research does not show

The record shows collapse under research conditions and leaves four questions open. A reader who sees only the headline papers should know which claims the record does not support.

  • No measurement of collapse in a deployed frontier model is on record.
  • No lab statement attributing a quality loss to model collapse is on record.
  • No measured share of model-generated content in web data is on record.
  • No price premium for human-made or older data is on record.

Each gap is a statement about the record, not about the world: a measurement may exist that the record does not hold, and the list changes only when a primary source is added.

The page does not take a side on whether these gaps will close. It records that, as of 2026-11-16, the evidence concerns research setups, and that statements such as "collapse is already happening" or "collapse is a myth" both go beyond it.

Why AI labs still generate training data despite collapse findings

AI labs generate training data with their own models and, in every disclosure on record, list it next to real data rather than instead of it.

Anthropic's Form S-1 of 2026-10-26 lists public web data, datasets licensed or commissioned from third parties, and synthetic data generated by its own models. The xAI summary for Grok 4.7 of 2026-09-21 answers "Yes" for synthetic data generated with its own models, for user data and for private third-party datasets.

Apple's 2025 foundation models report names responsible web crawling, licensed corpora and synthetic data. No lab states a synthetic share, so the disclosures show that generated data is used and say nothing about how much.

They also do not show any lab replacing real data with generated data, which is the condition under which the research reports collapse.

The need-by-need comparison and the full lab table belong to the page on synthetic vs real training data, and they are not repeated here.

What buyers' listings show about demand for human-made records

Buyers of company data state in their own listings that generated, AI-written or scripted records are not what they want. The table quotes four entries from Mercor's listings, as of 2026-11-09, ordered alphabetically by listing.

Mercor listing "Not a fit" entry Checked Source
AI Product Company synthetic or scripted traffic 2026-11-09 api.data.mercor.com
Digital Health Company role-played or scripted visits 2026-11-09 api.data.mercor.com
Medical Practice or Health System synthetic or sample charts 2026-11-09 api.data.mercor.com
Software Company code mostly written by AI tools 2026-11-09 api.data.mercor.com

Mercor's own listing text, quoted in part; other entries on each listing are not shown here.

These are Mercor's statements, checked on 2026-11-09, and they come from one buyer's listings, not from a market survey. Each listing also names other entries as not a fit, such as missing consent or records that cannot be stripped of personal information, and several of those concern rights and privacy, not generation.

The terms and listings of that buyer are in Mercor's data buyer profile. Mercor states no reason for these entries, and none is inferred here.

Human feedback is a second case of people-made data, and it is collected from people, not generated. Meta's Llama 2 paper of 2023-07-19 reports 1,418,091 Meta human preference comparisons and 27,540 supervised fine-tuning annotations.

Epoch AI estimates the effective stock of public human-generated text at about 300 trillion tokens (90 % interval 100 to 1,000 trillion), published 2024-06-06 and checked 2026-11-09, and estimates that it is fully used between 2026 and 2032 (80 % interval).

That is Epoch AI's estimate, and it concerns supply of text, not collapse. It is the reason the question of human-made data arises at all: if public text is used up, developers look to licensed records, commissioned work and generated data, and collapse research bears on the last of these.

What collapse research means for the value of human-made data, and its limits

Collapse research supports demand for human-made data only as far as the record goes: it shows a risk under research conditions, not a price. Records that can be shown to be human-made, dated and documented answer the concern the papers raise, and training data provenance covers how origin is documented.

For a holder, the value sits in records a generator cannot imitate: real workflows, human judgments and dated archives. That reading is consistent with the four listing entries above, which exclude records that look generated or staged.

A buyer that wants real traffic, real visits, real charts or code written by developers is asking for the one property a generator cannot supply, which is that the records happened.

The limits matter as much. The accumulation result of Gerstgrasser and colleagues means that generated data does not by itself make human data scarce.

No price premium for human-made data is on record, so the page does not say that human data is worth more because of collapse, and it does not treat a risk in a paper as a price.

A holder who reads the research as a reason to expect higher payment is reading something the record does not contain.

How much company data is worth is a separate question, answered on the page about what company data is worth to AI companies.

Licensing human-made records in light of collapse research: benefits, risks, downsides and alternatives

A holder of human-made records can license them, wait, license a limited sample, or not license at all, and collapse research changes none of these options by itself. The lists below set out the trade-offs for a holder weighing the question, and each item is limited to what the record shows.

Benefits

  • Buyers state in their listings that they do not want generated or scripted records, as the table above shows for one buyer.
  • Documented origin answers the concern that the Nature paper raises about generated training data.
  • Under a license, the holder keeps the originals, which a sale would not guarantee.

Risks

  • The research may shift: if the accumulation result holds, generated data matters less as a threat to the value of real records.
  • No price premium for human-made data is on record, so demand stated in listings is not a price.
  • A buyer's listing can change, and the check date on each entry is the only date it is known to be true.
  • The two collapse preprints are not peer-reviewed in the record, so a license decision should not rest on either of them.

Downsides

  • Documenting where records came from is work for the holder, and no cost figure is on record; the same documentation also serves other buyer checks, which is a separate subject.
  • Preparing records for a buyer takes time, with no figure on record either.

Alternatives

  • License a limited sample first.
  • License non-exclusively, so the same records stay available to others.
  • Keep the records for internal use.
  • Not licensing at all, which is a valid outcome; the decision is covered in whether to sell or license your data at all.

Where model collapse meets synthetic data, human data work and provenance

This page covers the collapse research and what it means for human-made data; the pages below cover the neighboring questions.

Every figure and date on this page was checked against its primary source, following how we source, date and check every figure. The page is information, not legal advice.

Frequently asked questions

Is model collapse real?

Model collapse is real under research conditions. Shumailov and colleagues (Nature, 2024-07-24) and Dohmatob and colleagues (arXiv, 2024-10-07) report it, and no measurement in a deployed frontier model is on record as of 2026-11-16. The research describes a risk in a loop, not an event in a product.

Does mixing in human data prevent model collapse?

The record disagrees. Gerstgrasser and colleagues report bounded test error when real and synthetic data accumulate, while Dohmatob and colleagues report collapse from 1 % synthetic data in a regression setting. Both are arXiv preprints with no venue in the record, and the setups differ, so the record does not settle the question.

Does model collapse make human-made data more valuable?

It supports demand for records that are verifiably human-made, and it sets no price. Mercor's listings, checked 2026-11-09, exclude synthetic or scripted records, and no price premium is on record. The research is a reason for buyers to ask about origin, not a number.

Which paper describes model collapse?

Shumailov and colleagues, in Nature 631, pages 755 to 759. It was published online on 2024-07-24, after a preprint of 2023-05-27 (arXiv 2305.17493). It is the peer-reviewed source, and the two preprints in the record are separate work.