← News

Models & toolsSources: arXiv · 16 Nov 2026

arXiv paper links model collapse to cultural evolution, casting fresh human data as a paid supply

A theoretical paper argues that knowledge loss in models trained on model output follows two mechanisms from cultural evolution, and that someone must keep paying to learn.

In plain words

When new AI models are trained on text produced by earlier models, they tend to lose knowledge over successive generations, starting with rare material. The paper's authors connect this to cultural evolution research, which distinguishes imperfect copying in a stable world from perfectly copied information going stale in a changing one. They also argue that outside information, such as fresh human data, must continually be paid for by someone.

What was announced

The authors frame model collapse as a single chain with two separately measurable parameters: in a fixed world, copies are lossy, echoing Henrich's account of skill loss in small populations, while in a changing world even perfect copies go stale, which they tie to Rogers' paradox. Existing collapse theory treats the supply of outside information, such as fresh human data, verifiers or curators, as given, whereas the paper argues cultural evolution offers a theory of that supply.

Tests on language-model lineages, including word trigrams, a character transformer and OPT-125m on various text corpora and a world of 2,000 fictional facts, reportedly matched predictions with no fitted parameters in the static case and with the change rate set rather than fitted in the dynamic one. Models inheriting weights kept more in static worlds but carried more outdated facts in changing ones. When data writers could query the model instead of finding out directly, the model's reliability ended up matching the cost of finding out, and the authors derive when checking remains a stable alternative.

What it does not show

  • The source is only the paper's abstract, so the experiments and closed-form results are not independently verifiable from this excerpt and have not been confirmed by peer review or external replication.

For security and IT teams

Scope
The reported experiments cover small models and text corpora, including OPT-125m, so findings may not extend to larger production models.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.