WildAI-monthly

License: CC BY-NC-SA 4.0. The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.

WildAI-monthly is the monthly measurement sample of How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text: 310,000 English web documents, 5,000 drawn at random from each of 62 crawl months from January 2021 to August 2026, each labeled by the Pangram AI-text detector as human-written, AI-generated or mixed. It is the data behind the paper's measurement of how much of the web is AI-generated (Figure 1, left).

Unlike the WildAI pool, no detector chose which documents are in this sample: within each month the draw is label-blind, so it can be used to measure the web month by month.

Composition

Source	Crawl months	Documents	GPT-2 tokens
fineweb	January 2021 – June 2025 (48)	240,000	178,090,615
common_crawl	July 2025 – August 2026 (14)	70,000	48,791,831
Total		310,000	226,882,446

Pangram labels: 285,105 Human, 3,901 Mixed, 20,994 AI. Six calendar months have no Common Crawl captures and therefore no rows: February–April 2022, July–August 2023 and January 2024.

How it was built

1. FineWeb months (January 2021 – June 2025, source = fineweb). For each year, a fixed random set of 260 files of Hugging Face FineWeb (HuggingFaceFW/fineweb), balanced across that year's crawls (about 20M documents per year), keeping documents of at least 200 characters. A document's month is its capture month (date[:7]), and 5,000 documents of each month were drawn at random.
2. Common Crawl months (July 2025 – August 2026, source = common_crawl). For each month's crawl, 32 WARC files drawn uniformly without replacement from the whole crawl were run through FineWeb's recipe (URL filter, Trafilatura extraction, English, repetition and quality filters; DataTrove 0.2.0, Trafilatura 1.8.1) without MinHash deduplication, and 5,000 of the resulting documents were drawn at random.
3. Labeling. Pangram 3.3.2 (public API, default model) labeled each document's text, up to its first 200,000 characters, as Human, Mixed or AI (pangram_label), with the fractions of the text classified AI, AI-assisted and human. These labels score the whole document; the WildAI pool's labels score up to three 512-token windows instead.
4. Topics and formats. WebOrganizer's topic and format classifiers (WebOrganizer/TopicClassifier, WebOrganizer/FormatClassifier) labeled each document from its URL and text, up to 8,192 tokens.

Measuring the web

The paper's monthly AI share is the share of a month's GPT-2 tokens that lie in documents Pangram labels AI (or, as a second measure, AI or Mixed):

from datasets import load_dataset

df = load_dataset("pangram/WildAI-monthly", split="train").select_columns(
    ["month", "token_count", "pangram_label"]).to_pandas()
tokens = df.pivot_table(index="month", columns="pangram_label", values="token_count", aggfunc="sum", fill_value=0)
ai_share = tokens["AI"] / tokens.sum(axis=1)
ai_or_mixed_share = (tokens["AI"] + tokens["Mixed"]) / tokens.sum(axis=1)

This reproduces the paper's monthly shares exactly for January 2021 – June 2025, and to within 0.01 percentage points from July 2025, where anonymization at release (below) changed the token counts of a few documents.

Fields

Field	Description
id	Common Crawl WARC-Record-ID of the page
text, url, date, dump	Extracted text, URL, capture time (WARC-Date) and crawl, e.g. CC-MAIN-2026-25
month	Capture month the document was drawn for (YYYY-MM)
source	fineweb (Hugging Face FineWeb) or common_crawl (FineWeb's recipe rerun on random WARC files)
warc_path	WARC file of the capture; empty for FineWeb rows (FineWeb's file_path, joined on id, gives it)
language_score	fastText English score from FineWeb's language filter
token_count	GPT-2 tokens of the released text
truncated	The released text is shorter than the page's extracted text (never, here)
pii_anonymized_after_labeling	E-mail or IP addresses were anonymized at release, after labeling
pangram_label, pangram_fraction_ai, pangram_fraction_ai_assisted, pangram_fraction_human, pangram_version	Pangram's label and fractions
topic, topic_score, format, format_score	WebOrganizer topic and format

Limitations

• Detector labels are not ground truth: Pangram makes errors, and Mixed documents are ambiguous.
• Each month is a random draw from the frame described above, not from the whole web. FineWeb months inherit FineWeb's filters and deduplication; Common Crawl months come from 32 WARC files per crawl and are not deduplicated. The source changes in July 2025, so compare months across that boundary with care.
• A document's month is its capture month, which is not when the page was written.

Personal information and removal requests

FineWeb rows went through FineWeb's anonymization of e-mail and IP addresses. Common Crawl rows were anonymized the same way at release; pii_anonymized_after_labeling marks the 2,379 rows that changed. Other personal information may remain. To request removal of content, contact the dataset maintainers through the Hugging Face discussion tab.

License

License: CC BY-NC-SA 4.0. The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license. The full text is in LICENSE. The text comes from FineWeb (ODC-By 1.0) and Common Crawl, and remains subject to Common Crawl's terms of use.

Citation

@article{russell2026wildai,
  title   = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
  author  = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
  journal = {arXiv preprint arXiv:2609.40295},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.40295}
}