A new preprint study from researchers at Imperial College London, Stanford University, and the Internet Archive has quantified what many users have observed anecdotally: the internet has an AI slop problem. The research, published in April 2026, found that approximately 35 percent of all new websites created between 2022 and 2025 are either AI-generated or AI-assisted.
To conduct their analysis, the research team used detection tools from Pangram Labs after testing four different approaches to AI detection. They tapped the Internet Archive's Wayback Machine to compile a representative sample of websites and tested six different theories about the characteristics of AI-generated content.
One of the study's most striking findings concerns tone. Using sentiment analysis to classify words as positive, neutral, or negative, researchers found that AI-generated or AI-assisted websites showed positive sentiment scores 107 percent higher than non-AI websites. The researchers view this spike in artificial cheerfulness as a symptom of what they call the sycophantic and overoptimistic nature of existing large language models. In other words, AI writing tools' tendency to cater to their human users has a spillover effect, making the overall tenor of online writing more saccharine.
The study also found evidence that AI is reducing ideological diversity online. AI websites scored roughly 33 percent higher on testing for semantic similarity than human-made websites, suggesting that the range of unique ideas and diverse viewpoints available online is shrinking.
However, several of the researchers' initial hypotheses were not supported by the evidence. They had suspected that the rise in AI-generated writing would lead to more misinformation, but their analysis found no evidence for this. They also expected AI writing to avoid linking to external sources and to be stylistically more generic than human writing. Neither of those theories held up under scrutiny.
Stanford researcher Maty Bohacek noted the surprise on his team: "Everyone on the team expected that to be true. But we just don't have significant evidence for that."
Before conducting their analysis, the research team commissioned a poll asking how people felt about AI. Comparing it to their findings, they discovered that public expectations about AI content often diverged from reality. Most respondents had assumed that more AI-generated websites would lead to a rise in fake news, that AI writing would stop linking to external sources, and that it would develop an increasingly generic voice. Like the researchers themselves, the public tended to expect the worst outcomes.
Bohacek emphasized that this study is a starting point rather than a definitive statement on AI's impact. "We just wanted to break ground," he said, positioning the research as a jumping-off point for deeper exploration into how AI is reshaping the internet.