If you own a smartphone and regularly browse the internet, you have almost certainly encountered what many call the internet’s “AI slop problem” — and it is impossible to miss. The crisis has grown dramatically worse since ChatGPT launched in 2022, with major social platforms now overwhelmed by AI-generated text of every kind. For years, this frustration was based on anecdotal experience alone; now, new research has put hard data behind what casual users have long observed.
A new preprint study, published by a collaborative team of researchers from Imperial College London, Stanford University, and the Internet Archive, finds that roughly 35 percent of all new websites are fully AI-generated or partially written with AI assistance. Beyond counting the volume of AI content, the study also confirms that online text is “increasingly sanitized and artificially cheerful.” Put simply: AI is making the entire internet feel fake-happy.
To reach their conclusions, the research team tested four distinct AI detection methodologies before settling on tools built by Pangram Labs, which delivered the most consistent, reliable results across tests. Even with that strong performance, the team notes a critical caveat: every AI detection tool currently available is imperfect, and false readings are always possible. To build a representative sample of public websites, the team pulled archived snapshots from the Internet Archive’s Wayback Machine, covering sites launched between 2022 and 2025. In addition to measuring how many of these sites relied on AI for their written content, the study also tested six widely held theories about the core traits of AI-generated slop.
One analysis focused specifically on the trend of artificial cheer, measuring how AI shifts the overall tone of online writing. Using sentiment analysis, which categorizes language as positive, neutral, or negative, the team found that the average positive sentiment score for AI-generated or AI-assisted sites was 107 percent higher than the score for fully human-written websites. Researchers describe this spike in unnatural positivity as a symptom of the “sycophantic and overoptimistic nature of existing large language models.” In short, AI writing tools’ default tendency to please their human users spills out into the broader internet, leaving the overall tone of online writing far more cloying and ungenuine.
A second test examined whether the boom in AI content has shrunk the range of unique ideas and diverse perspectives available online. This hypothesis was confirmed: AI does make the internet less ideologically diverse, with AI-written sites scoring roughly 33 percent higher on semantic similarity testing (a measure of how closely text matches existing content) than human-built websites.
While two of the team’s core assumptions about AI content held up to testing, the other four common hypotheses were not supported by evidence. Most notably, the team had predicted AI use would lead to a measurable rise in online misinformation, but their analysis found no data to back this widely held belief. They also guessed that AI writing would include fewer links to external sources than human writing, and that it would be far more stylistically generic than human work. Contrary to nearly universal expectations, neither of these assumptions proved true.
Even though the study confirmed AI content is more ideologically homogeneous and consistently upbeat, it found no evidence that AI writing as a whole has a flatter, more generic style. This outcome shocked the research team, which had entered the project expecting to see a clear shift toward generic, uniform output. “Everyone on the team expected that to be true,” says Stanford researcher Maty Bohacek. “But we just don’t have significant evidence for that.”
Before conducting their content analysis, the team commissioned a public opinion poll to measure how ordinary people perceive AI-generated online content. Comparing poll results to their empirical findings, the team discovered researchers were not the only ones whose expectations were upended: most commonly held beliefs about AI writing are wrong, the study concludes.
Just like the study’s authors, most poll respondents assumed that more AI-generated websites would mean a corresponding rise in fake news online. A vast majority also assumed AI writing would include fewer links to external sources, and that it would have an increasingly generic, uniform tone. “It’s interesting to see that people tended to expect the worst outcomes,” Bohacek says.
This research is far from the final word on AI’s impact on the internet. “We just wanted to break ground,” says Bohacek, who frames the work as a jumping-off point for deeper, more targeted exploration in the future. As a snapshot of AI slop’s early impact on the web, it offers a distinctly human takeaway: when it comes to new technology, predicting exactly how things will unfold is often far harder than we expect.
