A paper posted to arXiv on 30 September puts a number on something the field has been arguing about since 2023: how much of the open web is now written by machines, and what happens to a model trained on it. The answer to the first question is 27.5% of tokens in June 2026 web data, rising to 31.1% by August.
The answer to the second is more interesting, because it is not the collapse story. “How Much Is an AI Token Worth?” finds that AI-generated text is useful under some conditions and harmful under others, and identifies which.
What they actually did
The authors trained 800 language models across varying ratios of AI to human tokens, then measured loss on held-out human text. Two regimes fall out of the results.
For a model that is data-limited — not enough human text to go round — adding AI tokens initially reduces loss on human text. The benefit then saturates and reverses. For a model trained on a substantial quantity of human text, AI tokens increase loss immediately. There is no window in which they help.
That distinction matters commercially. A lab with licensed archives and a large human corpus gains nothing from scraped synthetic text and loses something by including it. A team training on whatever it can crawl is in the first regime, where some synthetic content is a net positive until it is not.
The scaling law
The paper proposes a revised scaling law that accounts for the AI-token ratio. The authors report it predicts held-out human-text loss for models up to 3.6 times larger with 41% lower error than the best existing law across all AI ratios. They also release WildAI, an 83-billion-token corpus with per-document labelling, alongside the 800 trained models.
The load-bearing assumption
Every figure here rests on the detector. The labels come from Pangram, a commercial AI-text classifier, and two of the paper’s authors — Max Spero and Bradley Emi — are Pangram Labs’ chief executive and chief technology officer. The paper is not hiding this; the affiliations are on it. But a measurement of how much of the web is AI-written is only as good as the tool deciding, and the tool here is sold by people on the author list.
What makes the result checkable anyway is the release. WildAI’s labels and the 800 models are public, so anyone with a different detector can relabel the corpus and see whether the scaling law survives.
What to watch
Whether an independent group reproduces the 27.5% figure with a different classifier. If the ratio holds, the data question stops being about availability and becomes about provenance: not how many tokens a lab can get, but how many of them were written by something else’s model.
