Researchers at the Pew Research Center conducted a study targeting approximately 500,000 websites created between 2021 and 2026. Using the AI detection tool "Open Pangram," they investigated whether the text contains traces of being created or significantly edited by generative AI.
Over One-Third of Websites Published After ChatGPT's Debut Show Signs of Generative AI Creation
This article is a translation. Read the Japanese original
The results of the study confirmed traces suggesting AI creation in 10% of the sample. When excluding sites created before the release of ChatGPT and focusing solely on sites published after its release, that proportion rises to more than one-third.
The percentage varies depending on the domain type. In ".com" domains, traces were found in approximately 10% of pages. This is about twice the 4.6% seen in ".org" domains, and nearly ten times the approximately 1% seen in ".edu" and ".gov" domains.
Characteristics specific to AI writing include the use of em dashes, Oxford commas, and certain negative expressions. Comparing 2023 to 2026, all of these features showed an increasing trend, with the appearance of negative expressions specifically shown to have increased threefold.
Source: Study finds traces of "AI" in one-third of websites published since ChatGPT's debut (GIGAZINE, 2026-09-06)