According to a new analysis from Pew Research, more than a third of web pages published since the launch of ChatGPT bear clear signs of AI writing or editing. A figure that speaks less of a technological revolution and more of a silent addiction, that of readers who delude themselves into still distinguishing a human text from an artificial one.
In the sample analyzed by Pew Research Center in July 2026, nearly half a million English-language web pages collected by the archive Common Crawlthe starting data seems almost reassuring: only 10% of the total shows significant signs of AI writing. But it is a misleading number, because it mixes very old pages, published well before ChatGPT existed, with very fresh content.
The real data comes when researchers have filtered out the past and look only at what was published after November 2022, the month in which OpenAI released ChatGPT to the public. In that subset, over a third of the pages35%, bear signs of authorship or artificial editing. It is no longer a niche phenomenon: it is the normality of the web that is being written today.
The traces that AI leaves without wanting to
What makes the research interesting is not just the percentage, but how it is measured. The survey model used by Pew, Open Pangramdoes not look for isolated “evidence”, but statistical patterns: words, punctuation, syntactic constructions that generative AI uses much more often than human authors.
Compared to 2023, the em dash have doubled. The Oxford commas they grew by 63%. Words like “crucial”, “fundamental”, “deepen”, “intertwining” have more than doubled. And the rhetorical construction “it’s not just X, it’s Y,” that little trick that gives an air of depth to any sentence, has almost tripled. Weak signals, taken individually, but which become a fingerprint when observed at scale.
There’s also some surprising geography. The domains .com show signs of AI writing ten times more often than domains .edu or .govwhile i .org they are halfway there. The commercial web, the one that thrives on traffic and publication speed, is also the most exposed.
The addiction that no one admits
The numbers give way to a more uncomfortable question, which concerns not so much the machines as us. If a third of new content bears the mark of AI and the majority of readers do not notice it, it means that we’re used to itor we are doing it, to a type of writing without realizing it. The serious thing is that this habit is eroding more than our ability to guess who wrote what.
It is eroding the critical thinking itself, that attention span that once led us to ask ourselves who is speaking, where information comes from, with what interest it was written. When a text sounds more and more “correct”, more and more fluid, more and more the same as itself, we stop asking that question. Readability becomes an anesthetic.
And the problem isn’t that the AI writes badly. It’s that he writes too goodin a way recognizable only to those who look for it on purpose, with tools like those used by Pew. The average reader doesn’t have a statistical detector in his pocket. He just has the habit, and the habit, faced with a constant flow of well-crafted and indistinguishable texts, stops asking questions.
A tool that can no longer be refused
It would be convenient, at this point, to dismiss everything as an alarm from techno-skeptics. It is not, and never will be for those who work in the media. Generative AI is now one co-working tool that no information operator can afford to ignore: it accelerates research, helps structure, suggests angles that would otherwise take hours. Rejecting it on principle would mean giving up a part of the profession that has already changed forever today.
The point is no longer “to use it or not”. AND as it is used, and above all what remains visible, declared, the boundary between what an editorial team signs with its editorial responsibility and what an algorithm produces without real human control.
The real risk is not the machine, it is our inattention
Pew data doesn’t say AI is replacing humans in writing. They say something more subtle and perhaps more urgent: that we are losing the ability, or perhaps the desire, to distinguish. And that this loss occurs just as AI becomes indispensable, not despite it.
The challenge for professional writers today is not to compete with AI on speed or fluidity, a losing game from the start. It’s staying recognizable: bringing into the text something that no linguistic model can truly replicate, a point of view, a direct experience, a doubt that cannot be resolved in a perfect sentence.
Perhaps this is precisely the most honest definition of authenticity today.
