Study finds no link between AI-generated text and factual errors online
A new study measuring actual web content found no statistical connection between the growth of AI-generated text and an increase in factual errors. The research contradicts widespread public concern about internet quality as ChatGPT and similar tools proliferate.
By mid-2025, roughly 35 percent of new web pages were created fully or partly with AI, up from zero before ChatGPT launched in late 2022. Researchers analyzed pages from the Internet Archive's Wayback Machine spanning August 2022 to May 2025 to measure whether content quality actually declined.
How researchers measured the impact
The study used the Pangram v3 detector to identify AI-generated text among web pages. Researchers then extracted concrete claims from those pages and had 50 people manually verify their accuracy. They also surveyed 853 U.S. adults about their beliefs regarding six potential harms from AI text.
The measurement showed no correlation between AI text volume and factual errors. If anything, the data leaned slightly toward fewer errors as AI content increased-but the relationship was not statistically significant.
Public fears didn't match measured reality
Three widespread concerns found no support in the data:
- 83 percent of survey respondents believed AI would erase individual writing styles, but measurements showed no link between AI adoption and stylistic uniformity.
- 70 percent feared articles would stop linking to external sources. The data showed no connection.
- 61 percent thought AI text would become longer while saying less. Again, no correlation appeared in the measurements.
The largest gap between perception and measurement involved factual accuracy. Three-quarters of respondents believed AI text would flood the internet with false information. The actual content analysis found the opposite trend, though researchers stopped short of claiming AI improved accuracy.
Two measurable differences emerged
AI-generated pages did show two distinct characteristics compared to human-written content. They contained 33 percent higher similarity between ideas and perspectives, suggesting less diversity in viewpoint. They also used 107 percent more positive language.
Experience shapes perception
Frequent AI tool users expressed less concern about negative effects than those who rarely use the technology. Among regular users, 76 percent believed in potential harms. Among infrequent users, 88 percent did. People with negative views of AI overall were most likely to accept claims about its dangers.
The disconnect between public worry and measured outcomes suggests that beliefs about AI text quality may reflect broader attitudes toward the technology rather than observable changes in web content.
Your membership also unlocks:
