The problem begins before training
AI laboratories have turned the internet into raw material. The operation once seemed to have a simple rule: gather more text, clean it and train on it. Now the material contains growing quantities of answers produced by earlier models. If a system learns from pages written by people and pages manufactured by systems, what exactly is it learning? More language, or an increasingly recycled version of itself?
The preprint *How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*, published on arXiv on 30 September 2026, puts numbers on part of the problem. Jenna Russell and her co-authors describe Wild AI, an extension of FineWeb that labels internet documents as human-written or AI-generated, then train 800 small models on different mixtures of the two. They compare how well each model predicts several evaluation sets. The paper asks a practical engineering and training-economics question: given a data budget, which mixture improves the chosen objective? The cultural value of machine prose lies outside the experiment.
The team estimates that documents labelled AI-generated account for 27.5 per cent of tokens in its June 2026 crawl and 31.1 per cent in August. It is tempting to turn those figures into the headline “a third of the internet is already AI”. The study does not measure the whole internet. To build Wild AI, an extension of FineWeb, the researchers apply quality filters and use EditLens to preselect likely AI-generated documents; Pangram then checks documents labelled with high confidence. The monthly series uses a different design: Pangram labels 5,000 documents for each month in the series for which a Common Crawl snapshot is available. The documents are sampled from material that passes FineWeb’s filters. The figure describes that procedure, sample and set of crawled pages; it cannot establish what proportion of everything published online was written by a machine.
That distinction changes the reach of the number. Pangram is part of the instrument with which the study classifies the web. The authors report its false-positive and false-negative rates from another paper, and run additional checks on pages from before ChatGPT. EditLens’s initial selection also matters. The paper says plainly that Wild AI is enriched by design with documents the first classifier considers likely to be artificial; its distribution does not match a natural sample of the internet. The result is useful for studying trends and training choices, alongside uncertainty introduced by the procedure itself.
The answer depends on what a model is meant to learn
The training experiment produces a less spectacular conclusion than the percentage, but a more useful one for laboratories. When the objective is to predict human text, adding AI-generated data helps when human material is scarce. As the amount of human text available per model parameter rises, the benefit disappears and can become a cost. On the C4 evaluation set, the share of synthetic text that minimises loss falls from 37 per cent when the human-data budget is low to about 1 per cent at twenty human tokens per parameter.
“Loss” measures how well a model predicts the next token; a lower value usually indicates a better fit to the evaluation set. That figure describes prediction, not reasoning, usefulness, safety or overall quality. The distinction matters because the evidence supports a limited claim: the useful mixture changes with the supply of human text and with the kind of text the model will be expected to process.
For evaluating AI-generated text, the answer reverses. Models improve when they receive much more synthetic material if they are also assessed against synthetic text. The symmetry makes sense: a system can learn efficiently the register it will later be asked to predict. It also exposes an evaluation trap. In one analysis, adding AI text raised loss on the human portion of a mixed dataset, while a validation set containing 22.3 per cent AI-labelled text reported 95.5 per cent of those harmful runs as improvements. When everything is blended into one number, the ease of predicting synthetic text can conceal a decline in predicting human writing.
The paper’s practical recommendation is clear: report the results separately. The authors would separate performance on human and synthetic text instead of allowing a combined validation set to speak for both. In a direct comparison, they trained 19.9-million- and 35.8-million-parameter models with twenty human tokens per parameter, then added data in successive passes: repeating the human corpus reduced loss on C4 more than adding new AI text. The findings on other evaluation sets are separate analyses; there the paper estimates which data mixture minimises loss for each target. The direct test is limited and does not establish a rule for every model size, language and task. Across the study, models range from 19.9 million to 973 million parameters, and the proposed law is fitted on models up to 268 million before being tested on the larger ones.
The paper extrapolates its findings and projects that, if AI text continues along the observed path, its share will reach 42.3 per cent by the end of 2027 and 50.7 per cent by the end of 2028. On that basis, the authors estimate that unfiltered training could require up to three times the compute of a recipe based on the human subset to match its loss on C4. These are scenarios produced by a scaling law and a time-series forecast, not measurements of the future. The study also covers only pre-training on English web text; its authors acknowledge that the models are too small to evaluate meaningful capabilities. Their conclusions about frontier-scale models still await a direct test.
The researchers disclose how much work AI did
The AI-use statement deserves close attention. The authors say coding agents proposed and tested forms of the mathematical law; suggested experiments, including the human-data repetition test; gave methodological feedback; designed and ran training for all additive models; implemented all experimental code; extended FineWeb with recent data; and produced tables and figures that made it into the manuscript. The team says it reviewed and validated the work and accepts responsibility for the text and results. It also discloses support and funding from Pangram Labs, the company whose detector is part of the study.
The disclosure shows what work was automated and which judgements the authors say they retained. Pangram appears both in the measurement chain and among the project’s funders; knowing that relationship helps readers interpret the detector’s role and makes public access to the data and code valuable. Wild AI and all 800 models are listed as available materials, allowing other teams to examine classification choices and repeat experiments.
The study’s merit is to bring an abstract concern—the possibility that AI will feed on its own output—down to budgets, evaluation sets and filtering decisions. Its most useful finding for laboratories is that a combined metric can register progress while a system worsens on one of the populations it must process. If a model must read human documents, researchers should measure its performance on them. If it will also consume synthetic text, that ability needs its own result.
Wild AI’s percentage will attract attention. The question that should follow it is more uncomfortable for the industry: what kind of language is each training set rewarding, and what disappears from view when a laboratory compresses it into a single score? The monthly estimate relies on Pangram labels applied to a filtered sample, not on a comprehensive manual classification of the web. The study itself offers laboratories an immediate measure: report loss on human text and AI-generated text separately. That shows which population improves when a new data mixture appears to raise overall performance.
Principal source
Jenna Russell et al., “How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text”, arXiv:2609.40295v1, 30 September 2026. Full text, methods and AI-use statement: HTML version. Data and code: WildAI repository.