The web is not just filling with AI text. It is learning one polite voice

A WEB THAT SOUNDS MORE ALIKE

The new analysis does not merely say that AI text has spread across the web. It says the web has become measurably more uniform and more cheerful because of it. That is less spectacular than panic over misinformation, but it may be a deeper long-term problem for public discourse.

According to The Decoder, researchers from Imperial College London, the Internet Archive, and Stanford analyzed a representative sample of English-language pages from the Wayback Machine. The sample covers 33 monthly intervals from August 2022 to May 2025, with AI text identified using the Pangram v3 detector.

The headline number is 35%. By May 2025, about 35% of newly published websites were fully or partially AI-generated, while the share was essentially zero before ChatGPT launched in late 2022. That is an industrial shift in less than three years.

The more important finding is the shape of the shift. AI texts were 33% more semantically similar to one another than human writing. In other words, automation is not only producing more content. It is producing content that more often clusters around the middle.

CHEERFULNESS AS SYSTEM TONE

The study's other supported effect is a positivity shift. AI texts scored 107% higher on positive sentiment than fully human-written content. That does not mean every AI text is funny or falsely upbeat, but at scale it shows a systematic pull toward pleasant, smooth, affirmative prose.

That fits the way models are tuned: helpful, polite, safe, and agreeable. The problem begins when that tone spills into millions of pages and starts to look like the neutral language of the web. Dissent, irony, sharpness, and local voice do not necessarily vanish, but they have to break through a thicker layer of generic pleasantness.

The study also matters for what it did not prove. The researchers did not find statistically solid evidence of more clearly false claims, nor did they confirm the disappearance of individual writing styles in the form they tested. They acknowledge limits: the analysis covered English text, detector reliability remains a dependency, and the truth-checking portion was much narrower than the main sample.

The best conclusion is therefore not that the internet has collapsed into AI garbage. It is narrower: the web may be filling not only with falsehoods, but with a tone that reduces difference between voices. For publishers, writers, and search systems, the next scarce value may be language that sounds less perfect and more real.