Adding AI-generated text scraped from the web can stop helping a language model—and may raise its loss on human-written text. A September 30, 2026 preprint reports that this turning point depends on how much human text the model has already seen, its size, the proportion of AI text added, and whether it is evaluated on human or AI-generated text. That is a conditional result about one kind of training data, not proof that all synthetic data is harmful or that the original Chinchilla result is broken.
What is the AI data satiation point?
It is the point at which adding more AI-generated web text no longer improves a model’s performance on held-out human text and can make it worse. The term describes a changing balance, not a universal token threshold: in the reported experiments it depended on model size, the existing human-text budget, and the ratio of AI to human tokens.
As an Amazon Associate I earn from qualifying purchases.
“Wild” AI text means unlabeled AI-generated writing that has entered ordinary web collections. It is different from synthetic examples deliberately generated and curated for a particular training task. It is also different from experiments in which successive models are recursively trained on earlier models’ outputs—the setup commonly discussed as model collapse.
What did the 2026 study find?
Jenna Russell, Ben Glickenhaus, Katherine Thai, John Wieting, Mohit Iyyer, Max Spero, and Bradley Emi report pretraining 800 language models with varying amounts of added AI-generated text. They fitted scaling laws to held-out loss on human and AI text. Their results suggest the value of added AI text changes with the starting conditions and the evaluation target.
#1 Best Overall
When human text is the target
For models with little human training data, some added AI text initially helped, but its benefit saturated and then reversed as more was added. For models with larger human-text budgets, adding AI text raised loss on held-out human text almost immediately in the reported experiments, while additional human text continued to lower it. The study therefore does not identify one ratio or threshold that applies to every model; the satiation point shifts with the regime.
When AI text is the target
The paper reports that AI-generated text can remain useful when the evaluation target is AI-generated text. That makes “AI text is harmful” too broad a conclusion. A model’s validation mix matters: a combined score can conceal a worsening result on its human-text slice if performance on AI text moves in the opposite direction.
The proposed scaling law
The authors propose a law with distinct terms for AI text’s modeled benefit and harm, allowing its marginal value to change sign. They report that the law reduces to Chinchilla when no AI text is included. When fitted on smaller models, it predicted held-out human-text loss for models up to 3.6 times larger with 41% lower error than the best existing law across the AI-text ratios they tested. Those are benchmark results within the paper’s experiments, not a guarantee for other corpora, architectures, or future frontier models.
Free tools Windows power users keep installed
One-click scans. No signup required.
Rank #2
Does this break the Chinchilla scaling law?
No. The new result challenges the use of a human-text-oriented scaling law to predict training outcomes after unlabeled AI-generated web text is mixed in; it does not refute Chinchilla’s original experiment or its conclusion under its own setup.
In their 2022 paper, Hoffmann and colleagues trained more than 400 language models, ranging from 70 million to over 16 billion parameters, on datasets from 5 billion to 500 billion tokens. Under their compute-optimal setup, they concluded that model size and training-token count should scale equally: double the model size and, correspondingly, double the training tokens. Their 70-billion-parameter Chinchilla model used four times as much training data as Gopher at the same compute budget.
The distinction is the data mixture. A relationship estimated for training on human text need not predict what happens when a changing share of the tokens is AI-generated. Nor is compute-optimal training the only objective: Sardana and colleagues’ 2024 inference-aware analysis argues that ordinary token-to-parameter ratios can overstate the impact of extra tokens at extreme ratios. That is a separate consideration about deployment economics, not evidence that the 2026 AI-text finding invalidates Chinchilla.
Rank #3
How much AI-generated text is in the web?
Russell and colleagues report that Pangram labeled 27.5% of tokens passing FineWeb quality filters in their June 2026 web sample as AI-generated; in their August 2026 sample, the labeled share was 31.1%. These figures describe the sampled crawl after those filters and the detector’s labels. They are not verified proportions of all web content, all online text, or the data used to train every model.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesDetection labels are measurements, not ground truth. A labeled corpus can help researchers examine provenance and compare data mixtures, but a label should not be treated as certain proof of how a passage was written. The authors report releasing WildAI, an 83-billion-token corpus with AI, topic, and format labels, along with models and code. Its current access conditions and license should be checked directly before reuse.
What should model builders do differently?
For a model whose goal is strong performance on human-written text, the authors recommend filtering AI text, repeating available human text before expanding the corpus with AI-generated web text, and reporting validation losses on human and AI text separately. This is the authors’ recommendation based on their results, not a universal standard for every training objective.
- State the target distribution. Say whether validation measures human text, AI-generated text, or a mixture; do not let a mixed aggregate stand in for both slices.
- Track the data mixture. Record the existing human-token budget and the AI-to-human token ratio so observed gains or losses can be interpreted in context.
- Distinguish corpus types. Unlabeled web text, task-specific curated synthetic examples, and recursive training on model outputs are different conditions and should not be treated as interchangeable.
- Test filtering and repetition against the objective. The paper’s result favors human text for a human-text target, but it does not establish that filtering is optimal when AI text itself is the target.
Will AI run out of human training data?
A 2024 position paper by Pablo Villalobos, Anson Ho, Jaime Sevilla, Tamay Besiroglu, Lennart Heim, and Marius Hobbhahn conditionally forecast that training datasets could approach their estimated stock of public human-generated text between 2026 and 2032 if then-current trends continued, or sooner under overtraining. This is a forecast based on assumptions, not a measured exhaustion date or prediction of inevitable model collapse.
The possibility of tighter public-text supply helps explain interest in synthetic data, but it does not settle which forms of synthetic data work, for which evaluation target, or at what mixture. A separate 2025 SynthLLM preprint reports a performance plateau near 300 billion tokens for its own synthetic-data framework and experiments. That result concerns a different setup and does not independently confirm the wild-web-text finding.
Quick Recap
What the result does—and does not—establish
- It does establish, as a reported preprint result: in the authors’ experiments, added wild AI web text had a conditional relationship with held-out human-text loss, including saturation and reversal in some regimes.
- It does not establish: a universal harmful percentage of AI text, that all synthetic training data degrades models, that every web corpus has the measured label rates, or that Chinchilla’s original compute-optimal result is false.
- Its practical implication: data provenance and the evaluation target matter. Report human- and AI-text validation separately before treating an aggregate score as evidence that either target improved.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
