According to TechCrunch, a study released by the Pew Research Center on the 20th local time found that more than one-third of web pages published since ChatGPT was launched in November 2022 may show signs of AI generation. The finding is consistent with other research, indicating that a large number of web pages added in recent years were either generated directly by AI or heavily modified by AI.
Shortly before the study was released, internet infrastructure services provider Cloudflare said that bot-generated web traffic had already exceeded traffic from human visitors, arriving earlier than previously predicted.
The Pew Research Center’s study focuses on which content on the internet is being accessed. The findings suggest that a large number of web pages may be forming a cycle in which bots read content generated by bots.
For the study, the Pew Research Center used the Common Crawl web archive to collect nearly 500,000 English-language web pages from approximately the past 5 years. Researchers then used technology from Open Pangram to determine whether the pages may have been written by AI or extensively modified by AI.
Among a randomly selected sample of 10,000 web pages in July 2026, about 10% showed clear signs of AI authorship.
The Pew Research Center noted that the random sample necessarily included some web pages published earlier, when AI writing tools had not yet emerged, meaning those pages could not have been generated by AI.
Therefore, to more accurately assess AI’s impact on internet content, the Pew Research Center further excluded older web pages published before ChatGPT’s launch and analyzed only content added after ChatGPT was introduced.
In this adjusted sample, 35% of web pages showed signs of AI authorship. By domain type, the proportion of .com websites showing signs of AI authorship was approximately 10 times that of .edu and .gov websites. In the latter two categories, the proportion of AI-created content was about 1% in each. The proportion for .org websites was 4.6%.
The Pew Research Center acknowledged that the analysis has limitations. Like other AI detection tools, Pangram may produce false positives and identify human-written web pages as AI-generated. However, the results still reflect the overall trend when viewed across a large sample.
The study also found that some expressions typically regarded as characteristics of AI writing have become more common in recent years: em dashes, Oxford commas, and sentence structures such as “This is not X, but Y.”
Note: The Oxford comma, also called the serial comma, is the comma used in English writing between the penultimate item and a conjunction such as “and” or “or” in a list containing three or more items. For example, in “A, B, and C,” the comma after B is the Oxford comma.