[Digital Today reporter Seung-a Yoo] A study has found signs that AI may have written or substantially edited content on 35% of webpages posted since ChatGPT was released.

On Aug. 25, IT outlet TechRadar reported that Pew Research Center analysed 10,000 webpages using data from the Common Crawl project. It said it found signs in 9.6% of the overall sample consistent with AI writing or substantial editing.

The study was based on English-language webpages collected from the Common Crawl web archive. Pew Research Center analysed about 490,000 webpages collected from 2021 to July 2026. It examined AI authorship based on a randomly selected sample of 10,000 pages from July 2026. The sample included old pages accumulated since the early days of the World Wide Web, meaning the overall share also included many pages written before AI appeared.

The share of the total sample showing signs of AI writing was 9.6%. Excluding old pages and looking only at pages posted after ChatGPT launched in November 2022, the share rose to 35%. Pew Research Center said this suggests the proportion of recent webpages involving AI in writing or editing is growing rapidly.

Differences by domain were also confirmed. In the 2026 sample, pages showing signs of AI writing were most concentrated in commercial .com domains. By contrast, .edu domains used by U.S. educational institutions and .gov domains used by government institutions were about 1% and 0.8%, respectively. The .org domain was about 4.6%. This can be interpreted as AI use being relatively more active on commercial sites than on education and government sites.

Pew Research Center also analysed about 490,000 English-language webpages to identify common linguistic traits seen in AI-written text. On pages posted since 2023, the frequency of the long dash known as the em dash doubled, and use of the Oxford comma rose 63%.

Use of certain words it said were favoured by AI also increased. The frequency of some words, including "delve", "interplay" and "testament", more than doubled after 2023. The frequency of negative parallel structures such as "not only X but also Y" also nearly tripled over the same period.

Pew Research Center stressed that these traits alone cannot definitively determine that a given text was written by AI. It said em dashes and the Oxford comma are also commonly used in human writing. Researchers said that rather than focusing on specific expressions in individual documents, a more accurate judgement requires combining statistical patterns in word choice and sentence structure seen across large-scale documents.

The study also shows the limits of AI content detection technology. It is difficult to distinguish human writing from AI-written or AI-edited text using only certain punctuation marks or words. Still, analysing large numbers of webpages can confirm a trend of AI-specific linguistic patterns increasing across web content.

It was also notable that the share of pages showing signs of AI writing varied widely depending on the nature of the website. Commercial sites showed relatively more traces of AI use, while education and government sites remained at low levels. As it becomes increasingly difficult to distinguish AI-generated content from human-written content, how to set standards for judging content credibility is emerging as a new challenge.