More than one in three English-language web pages published since ChatGPT's public debut show signs of being written or substantially edited by artificial intelligence, according to a new Pew Research Center analysis. The study, which examined roughly 490,000 English-language web pages collected between January 2021 and July 2026, found that 35% of pages created after ChatGPT's November 2022 launch exhibited indicators of AI authorship.
Across all pages in the dataset - including older content that predates AI writing tools - the figure stood at about 10%. Pew used the AI detection model Open Pangram to identify writing patterns typical of AI-generated text. The researchers acknowledged that the tool can misclassify individual pages, but argued that large-scale pattern analysis reveals meaningful differences in how humans and AI write.
Commercial sites lead, universities and government lag
AI writing traces varied sharply across domain types. Commercial .com sites showed an AI authorship rate of about 10%, more than double the 4.6% rate on .org domains. University (.edu) and government (.gov) domains remained near 1% throughout the five-year period.
- .com: 10.0%
- .org: 4.6%
- .edu: 1.0%
- .gov: 1.0%
The gap likely reflects different incentives: commercial sites have stronger reasons to produce high volumes of content quickly, while academic and government institutions face stricter editorial standards.
AI writing patterns spread across the web
Language habits favored by AI have also become markedly more common. Compared with 2023, em-dash usage roughly doubled, while Oxford comma usage rose 63%. Formal vocabulary preferred by AI - such as "delve," "interplay," and "testament" - more than doubled in frequency. Contrastive sentence structures of the "not simply X, but Y" variety nearly tripled.
Pew stressed that the numbers don't only capture text written entirely by AI. The figures include cases where AI substantially revised or polished human-written content. "Just because a person uses an em-dash or an Oxford comma does not mean their writing was produced by AI," the researchers said, adding that "when analyzing vast collections of documents together, we can observe meaningful differences in how frequently humans versus AI employ these expressions."
The findings carry direct implications for how AI systems are trained. As AI-generated content spreads across the web, the next generation of AI models will increasingly train on AI-written text rather than human-authored material. That raises concerns about "model collapse," where quality progressively degrades - akin to making copies of copies. For writers, this means the pool of genuinely human-written reference text is shrinking relative to machine output.
Why this matters for writers
If you're a working writer, these numbers describe your competitive environment. A third of new English-language web content now carries AI fingerprints, which means editors and readers are already encountering machine-generated prose at scale - and they're not always able to tell the difference. That doesn't make human writing obsolete; it makes the case for what you bring to the table: original reporting, verifiable facts, and a voice that doesn't follow the same statistical patterns as a language model.
It also matters for how you position your own work. If you're publishing online, the same detection tools Pew used can flag your writing if it leans on AI-favored conventions like heavy em-dash use or formal vocabulary. That's not a reason to avoid those features entirely - but it's worth knowing they carry a cost in credibility. For writers who want to build a durable audience, the practical takeaway is to write in a way that's clearly, demonstrably human: specific, sourced, and grounded in direct experience rather than polished generalities.
For those adapting to this shift, resources like AI for Writers offer practical guidance on working alongside AI tools without losing editorial judgment. Bloggers facing a web where AI content is the default may find the AI Learning Path for Bloggers useful for understanding how to maintain a distinct voice in a crowded field.
Separate data from Cloudflare shows bot traffic has already surpassed human web traffic, suggesting automated systems are shaping both the publishing and browsing sides of the internet. Pew's findings indicate that trend is visible in the text itself.
Your membership also unlocks:
