Pew Research Center found that AI-written content on the web appears with significant signs of AI authorship on about one in 10 English-language webpages. It reported that since ChatGPT’s launch in November 2022 the share rises to 35% of pages. Researchers analyzed roughly 490,000 pages from the Common Crawl web archive covering January 2021 through July 2026, testing that dataset to produce the estimates reported by the research center while the analysis examined English-language pages and quantified the proportions that the report directly presents as the study’s official headline findings.
AI-authorship rates vary by domain type: .com domains show approximately ten times higher AI-authorship rates than .edu or .gov domains, which are reported at about 1%, while .org domains are reported at about 4.6%. In January 2021, the AI-authorship rates across the four domain types were reported to be nearly identical. The reported proportions describe relative frequencies of substantial signs of AI authorship across those domain categories as presented in the findings.
The reported timeline for .com domains shows an increase from about 1% in January 2021 to 9.35% in January 2026. The figures reflect the change in measured AI-authorship rates on .com domains during the covered period. The paragraphs above summarize the reported domain-type differences and the reported timeline change for .com as presented in the findings.
Researchers tested webpages using Open Pangram, a detection tool from Pangram Labs. The tool was applied to identify substantial signs of AI authorship in web page text. Detection models can misclassify content, a limitation acknowledged by the researchers. The presence of significant signs of AI authorship on a page does not necessarily indicate that the page was entirely written by AI.
The researchers reported that academic and government sites tend to maintain editorial review processes. They reported that .com sites include newsrooms and content farms that publish faster than editors can review. The researchers also noted that AI fingerprinting may become easier if large AI companies introduce model-level text fingerprints, with Anthropic cited as an example.
These paragraphs summarize the researchers’ stated detection approach and the limitations they reported. They do not quantify prevalence or provide detection accuracy figures. The account above confines itself to the methodology and caveats described by the researchers.
Other research finds that roughly 9% of U.S. newspaper articles this year are AI-generated, including opinion pages of The New York Times. The report presents that figure alongside the web analysis. The 9% estimate refers to U.S. newspaper articles during the same year covered by the related research.
Detection of AI authorship involves analyzing stylistic elements such as em dashes and the Oxford comma as part of the analytical process. The researchers noted that detection models can misclassify content and that finding significant signs of AI authorship does not necessarily mean a page was entirely written by AI. The researchers also reported that academic and government sites tend to have editorial review while .com sites include newsrooms and content farms that publish faster than editors can review.
Merriam-Webster named ‘slop’ its word of the year. These related observations were presented alongside the study’s headline findings.
The study’s headline findings indicate that AI-written content on the web appears at varying rates across different domain types, with commercial domains showing substantially higher prevalence than academic or government sites and nonprofit domains occupying an intermediate position. Researchers reported that editorial oversight correlates with lower measured rates on academic and government sites, while commercial sites include newsrooms and content farms that publish faster than editors can review.
The research team used automated detection tools and noted limitations of those methods, including possible misclassification and that detected signs of AI authorship do not necessarily mean a page was entirely produced by AI. The researchers also noted that detection capabilities may evolve if major AI firms implement model-level text fingerprints, but current methods retain these stated caveats.
