How much of the new web is written by AI?
An independent, longitudinal dataset of AI-generated content in English-language news on the public web. Everything here is a population-level estimate with a confidence range, never a verdict on a single document. Data as of (UTC) · methodology v1.1.
Word-shift index
Baseline collecting
Needs 20 pre-2022-11 weeks; 0 so far
Leftover-phrase floor
Baseline collecting
A floor: documents with unmistakable chatbot leftovers
Baseline documents
0
0 of 20 weeks needed, English news before 2022-11-01
Population
English
News sites in Common Crawl CC-NEWS; other languages in a later version
What this page will show
No numbers are shown until there is a real baseline. We are collecting English news from before ChatGPT (2022-11) to compare against.
- Word-shift index: how much more often a documented list of words (10 style words from published studies) appears than in the pre-2022-11 baseline, with a 95% range. 1.00 means no change.
- Leftover-phrase floor: the share of documents that contain text such as “As an AI language model”. It is a minimum: most AI text has no such trace.
- Week by week for both, once the baseline holds 20 weeks and 10000 documents.
Methodology and version
Methodology v1.1. Documents come from Common Crawl CC-NEWS, sampled at random and kept with the exact location of the raw record. Estimates are population-level only; we do not label individual documents. Read the method.