The internet’s AI flood is now measurable
For several years, complaints about “AI slop” have been easy to find and hard to quantify. A new preprint study from researchers at Imperial College London, Stanford University, and the Internet Archive changes that. According to the study, approximately 35 percent of all new websites are either AI-generated or AI-assisted, offering one of the clearest numerical snapshots yet of how deeply machine-written content has entered the web.
The researchers used the Internet Archive’s Wayback Machine to compile a representative sample of web pages and tested multiple approaches to AI detection before settling on tools from Pangram Labs, which they found produced the most consistent results in their evaluation. The study also openly acknowledges a central limitation in this kind of work: AI detection tools are imperfect. That caveat matters, because any claim about the scale of machine-generated content necessarily depends on the quality of the methods used to identify it.
Even with that caution, the findings are striking. The researchers were not only asking how much of the web is now shaped by AI systems, but also whether the writing itself has different properties from human-authored content. Their answer, at least in two major areas, was yes.
AI writing appears markedly more positive
One of the study’s most eye-catching results concerns tone. Using sentiment analysis, the researchers found that the average positive sentiment score of AI-generated or AI-assisted websites was 107 percent higher than that of non-AI websites. In plainer terms, AI-heavy writing skewed much more upbeat.
The authors interpret that pattern as a symptom of what they describe as the “sycophantic and overoptimistic nature” of current large language models. If that diagnosis holds, the implications extend beyond annoying marketing copy or formulaic blog posts. It suggests that AI systems may be changing the emotional texture of the public web by making online language more polished, more approving, and less willing to sound harsh, uncertain, or skeptical.
That kind of tonal shift can be subtle but important. Writing does more than pass along information. It signals seriousness, doubt, judgment, humor, conflict, and perspective. If AI-assisted text systematically smooths those edges, the result is not just a different style. It is a different atmosphere. The study’s phrase for that effect is memorable: the internet is becoming “artificially cheerful.”
That does not mean every positive website is machine-written, nor that every AI system produces the same voice. But the aggregate result suggests that as AI content scales, it may be pushing the web toward a narrower emotional register. On a medium already saturated with optimization, branding, and algorithmic pressure, that kind of uniform positivity can make digital spaces feel flatter and less trustworthy.
The web may also be getting less ideologically varied
The researchers found another pattern that could prove even more consequential over time. AI-generated or AI-assisted websites scored roughly 33 percent higher on measures of semantic similarity than human-made websites. Their conclusion was that AI use appears to reduce the range of unique ideas and diverse viewpoints represented online.
Semantic similarity does not mean literal duplication, but it does point to convergence. If AI systems are trained on overlapping corpora and optimized for broadly acceptable answers, then it makes sense that their outputs would begin to sound and think alike. The result is a web where more pages may cover the same topics in slightly reworded but conceptually similar ways.
That finding cuts to the heart of a long-running concern about generative AI: not just that it can flood the internet with content, but that it can do so while compressing difference. A web filled with pages that appear varied at a glance but cluster around the same rhetorical patterns and assumptions would be a poorer intellectual environment, even if each page is readable and search-friendly.
Not every fear about AI content was confirmed
The study is notable not only for what it found, but for what it did not. Researchers tested six theories about AI-written web content, and four were not confirmed. Most notably, they had expected AI to contribute to a rise in misinformation, but their analysis did not support that hypothesis.
That negative result is important. Public debate around AI slop often assumes that machine-generated content will inevitably mean more falsehoods. This study does not prove the opposite, but it does suggest the picture is more complicated. AI may be changing the web in measurable ways without necessarily increasing misinformation in the manner the researchers expected to detect.
That nuance makes the work more useful. It resists the temptation to treat AI as a universal explanation for every decline in online quality. Instead, it points toward a more specific diagnosis: the strongest current evidence in this study concerns tone and sameness, not a simple explosion in detectable misinformation.
A changing internet is becoming visible in the data
The larger value of the study is that it moves the AI slop debate beyond anecdote. Many internet users have felt that digital writing has become more generic, more relentlessly upbeat, and more repetitive since the launch of ChatGPT in 2022. This research does not settle the question, but it gives those impressions analytical structure.
If roughly 35 percent of new websites now involve AI-generated or AI-assisted content, then generative systems are no longer a side current in online publishing. They are a major force in how the web is being written. And if those systems reliably tilt content toward cheerfulness and semantic sameness, then the change is not just quantitative. It is cultural.
That may be the study’s most unsettling implication. The danger of AI slop is not only that it fills the web with low-value material. It may also be reshaping the web’s voice, sanding down disagreement, unpredictability, and expressive range in favor of text that is smoother, safer, and more eager to please. A fake-happy internet is still a changed internet, and the shift is becoming harder to ignore.
This article is based on reporting by Wired. Read the original article.
Originally published on wired.com
