Edition of Sunday 15 November 2026

The Curse of Recursion: Training on Generated Data Makes Models Forget (2024)

Comments on “The Curse of Recursion: Training on Generated Data Makes Models Forget (2024)”

Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked

Researchers at Cambridge, Oxford, Imperial, Toronto and Edinburgh show that models trained recursively on their predecessors' output lose the tails of the original distribution and drift toward a narrow point estimate. They demonstrate this in Gaussian mixtures, variational autoencoders and fine-tuned language models, and argue from finite sampling error that it is unavoidable. For teams that train on scraped web data, human-generated data and provenance become more valuable. The language-model results come from fine-tuning, where only early signs appear, and several quoted figures lacked sources.

Read the original at arxiv.org. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Why it ranked

No. 4 of 62 with a score of 0.72. The panel rated it highest on technical depth (0.86) and audience breadth (0.85), lowest on new information (0.43). The judges disagreed most on new information (spread 0.70), so the story is marked contested.

AI judgeJudge 2claude-haiku-5-5
Widely discussed peer-reviewed-style study on model collapse from recursive training on synthetic data; strong evidence and high relevance to AI builders, though it is an older 2024 paper not among recent stories.
AI judgeJudge 3gpt-5.6-luna
A landmark foundational study introducing model collapse, with mathematical analysis and experiments across generative models. It is highly relevant to AI data pipelines, though the provided text predates recent follow-up work and includes claims that may be broader than demonstrated.

Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.

Score

0.72total under rubric v2.3Contestedjudges differ by 0.70 on new information

Judges’ mean answer per question, on a scale from 0 to 1. Questions marked minus count against the score.
  • Significance0.81
  • Reader interest0.83
  • Audience breadth0.85
  • Practical relevance0.64
  • Substantiated0.84
  • Technical depth0.86
  • New information0.43
  • Counts against: Promotional0.03
  • Counts against: Overclaim0.46

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

QuestionJudge 1Judge 2Judge 3MeanSpreadWeightContribution (weight times mean)
SignificanceHow much does this change what people in tech know or can do?0.740.800.900.810.16+0.24+0.195
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.700.850.950.830.25+0.26+0.217
Audience breadthHow many people who build technology does it concern?0.910.800.850.850.11+0.13+0.111
Practical relevanceHow useful is it to someone who builds software or tech products?0.580.600.750.640.17+0.22+0.142
SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations.0.880.800.850.840.08+0.05+0.042
Technical depthHow technically substantive is it?0.930.700.950.860.25+0.05+0.043
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.900.200.200.430.70 , above the contested threshold+0.05+0.022
PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional.0.060.020.000.030.06−0.19−0.005
OverclaimDo the headline or conclusions claim more than the body supports?0.630.300.450.460.33−0.11−0.051
Sum of contributions+0.715
Claim-check penalty0.000
Total, as published0.715

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.

The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.

Ranked #4 on Rubric, 15 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-15/4/"><img src="https://rubric.hrih.si/badge/2026-11-15/4.svg" alt="Ranked #4 on Rubric, 15 Nov 2026" width="238" height="28"></a>

Claim check

Claim check: 6 no source given · 2 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.

  1. No source givenCausal conclusion

    Use of model-generated content in training causes irreversible defects in the resulting models, where tails of the original content distribution disappear (termed model collapse).

    The article gives no link or citation for it.

    The article says

    “We find that use of model-generated content in training causes irreversible defects in the resulting models, where tails of the original content distribution disappear”

    The source says

    No source to compare with.

  2. No source givenGeneralization

    Model collapse can occur in Variational Autoencoders, Gaussian Mixture Models and LLMs.

    The article gives no link or citation for it.

    The article says

    “show that it can occur in Variational Autoencoders, Gaussian Mixture Models and LLMs”

    The source says

    No source to compare with.

  3. No source givenBenchmark result

    The fine-tuned OPT-125m model on wikitext2 reaches 34 mean perplexity from a zero-shot baseline of 115.

    The article gives no link or citation for it.

    The article says

    “The original model fine-tuned with real wikitext2 data gets $34$ mean perplexity, from the zero-shot baseline of $115$”

    The source says

    No source to compare with.

  4. No source givenBenchmark result

    With no original data preserved, training on generated data loses 20 to 28 perplexity points of performance on the original task.

    The article gives no link or citation for it.

    The article says

    “losing some performance – from $20$ to $28$ perplexity points”

    The source says

    No source to compare with.

  5. No source givenBenchmark result

    Preserving 10% of the original data leads to only minor performance degradation when fine-tuning on generated data.

    The article gives no link or citation for it.

    The article says

    “preservation of the original data allows for better model fine-tuning and leads to only minor degradation of performance”

    The source says

    No source to compare with.

  6. No source givenNumber

    A repetition penalty of 2.0 causes perplexity to double compared to the original, and models remain as susceptible to model collapse, if not more.

    The article gives no link or citation for it.

    The article says

    “enforcing this for the LLM experiments causes the perplexity to double, compared to the original. Models remain as susceptible to Model Collapse, if not more”

    The source says

    No source to compare with.

  7. Couldn’t checkCausal conclusion

    Google downgraded farmed articles in search results, favouring trustworthy sources such as education domains.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    No quote stored.

    The source says

    No passage could be compared.

    Source cited by the article: googleblog.blogspot.com/2011/02/finding-more-high-quality-sites-in.html

  8. Couldn’t checkCausal conclusion

    DuckDuckGo removed content-farm articles altogether.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “while DuckDuckGo removed them altogether”

    The source says

    No passage could be compared.

    Source cited by the article: technologyreview.com/2010/07/26/26327/the-search-engine-backlash-against-content-mills

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN 2 pts, not on front page

Discussion on Hacker News

Recorded for context only, checked 15 Nov 2026, 20:46 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition. How ranking works describes rubric v2.3.

Comments

Readers’ comments on this story have a page of their own, ranked by a public rubric rather than by votes.

Comments on “The Curse of Recursion: Training on Generated Data Makes Models Forget (2024)”

More on Rubric

From the last 7 editions, chosen by shared topic, score and date.