Pangram-backed study: AI-labelled web text helps data-starved models but raises loss on well-fed ones

A preprint posted to arXiv on 30 September 2026, "How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text", reports that after FineWeb quality filtering, 27.5% of June 2026 web tokens and 31.1% of August 2026 tokens are labelled as AI-generated by Pangram's detector.

An interest to note. Four of the seven authors are shown with a Pangram Labs affiliation, and the lead author, shown at the University of Maryland, did the work "during an internship at Pangram Labs". Pangram sells AI-text detection, and the acknowledgements thank its team "for the support and funding for this project". The web AI shares in the paper come from Pangram's own detector. The paper is a preprint; its arXiv listing carries no comments field or journal reference.

What the authors measured

• The web share. The authors ran Pangram 3.3.2 on 5,000 documents per crawl month. The share of FineWeb tokens labelled as AI-generated by Pangram's detector was 10.1% in June 2024, 16.1% in June 2025 and 27.5% in June 2026, they report.
• Detector error. The paper cites "a reported false-positive rate of 0.05% and false-negative rate of 1.99%" for Pangram 3.3.2, from a Pangram technical report whose authors overlap with this paper's. As its own check, Pangram 3.3.2 labelled 37 of 60,000 pre-ChatGPT documents from 2021 as AI (0.062%), which the authors treat as an upper bound on the false-positive rate.
• Quality filters. On a 10,000-document 2026 sample, AI-labelled documents passed FineWeb's pipeline 2.3 times as often as human-labelled ones, and DCLM's 9.8 times as often, the authors report.

What the training runs showed

The 800 models range from 19.9M to 973M parameters. In the abstract's words: "For data-starved models, adding AI tokens to pretraining data initially lowers loss on human text, but the benefit saturates as more are added and quickly reverses into harm. For models trained on high budgets of human text, AI tokens raise loss almost immediately, while the same number of fresh human tokens keeps lowering it."

The paper says AI text "lowers loss on human text below about 10 human tokens per parameter, but from the Chinchilla-optimal 20 its benefit disappears". On a 2026 mix with 22.3% AI-labelled text, models above about 10 to 15 tokens per parameter (fewer at larger sizes) had lower loss on C4, a human-text set, once the AI documents were removed.

The proposed scaling law

The authors say existing laws such as Chinchilla, "which treat AI tokens as no different from human tokens, fail to predict this behavior". Their law adds separate benefit and harm terms that let "the value of an AI token (measured in human token equivalents) ... change sign" and reduces to Chinchilla when there is no AI text.

They fit it on 726 models of 19.9M to 268M parameters and tested it on 74 models of 477M and 973M. The paper states that, fitted on smaller models, it "predicts the effect of AI text on held-out human-text loss for models up to 3.6x larger with 41% lower error than the best existing law over all AI ratios".

From that law the authors estimate that, at 268M parameters and 20 tokens per parameter, training on unfiltered web text at August 2026's share labelled as AI-generated by Pangram's detector needs 1.6 times the compute of training on its human subset. Projecting forward from those detector labels, they forecast that "over half (50.7%) of FineWeb tokens will be AI-generated by the end of 2028" (80% interval: 38.0% to 63.4%).

Caveats the paper itself raises

• The study measures next-token loss. The authors write that "at our sizes AI text raises CORE scores ... about as much as fresh human text does", while adding that "all models are too small to have meaningful performance, and these are the tasks AI-generated text was meant to optimize".
• It covers "only pretraining, on English web text labeled by Pangram".
• On model collapse, the authors distinguish their setting: "Unlike synthetic data or model-collapse setups, this wild AI text comes from many models, is written for human readers, and arrives unlabeled in pretraining corpora." Our earlier explainer, Model Collapse: What Happens When AI Trains on AI-Generated Data, argued the collapse scare is "largely avoidable".

What is released

The authors release WildAI, an 83B-token labelled corpus, all 800 models and code. The repository's LICENSE file and the Hugging Face dataset and model cards give CC BY-NC-SA 4.0, a non-commercial licence; the README says the dataset's text also remains subject to FineWeb's ODC-By 1.0 licence.

Why it matters

The authors recommend "filtering AI text when the target is human text", repeating human text before adding AI web text, and reporting validation loss on human and AI text separately, since at a 22.3% AI-labelled share a mixed validation set "hides the harm in 95.5% of harmful runs". On our reading, the results rest on one vendor's detector and on models below 1B parameters, both of which the paper acknowledges.