The word ‘delve’ has more than doubled in usage on the internet since 2023. So has ‘tapestry’, ‘testament’, and ‘interplay’. These aren’t vocabulary trends driven by human writers discovering new favorites. They’re statistical fingerprints, and they point to a single source. According to Pew Research Center, over one-third of English-language webpages published after the release of ChatGPT now show significant signs of AI authorship. That’s not a rounding error. That’s a structural shift in how the web is produced.
What Pew actually studied
The research team pulled almost 490,000 English-language webpages from the Common Crawl web archive, spanning the five years from before ChatGPT’s November 2022 launch through July 2026. They then ran each page through Open Pangram, a machine learning-based AI detection tool that looks for linguistic patterns statistically more common in AI-generated text than in human writing.
The headline number is 10% of all pages sampled in July 2026 show significant signs of AI authorship. But that figure includes enormous amounts of older content that simply couldn’t have been AI-written. Filter for pages published after ChatGPT’s release, and the number jumps to over 33%. That’s the more honest benchmark for what the current web looks like.
Where AI text concentrates
The distribution across domains tells its own story. Commercial .com domains show AI authorship signs at roughly 9.4%, about double the rate on .org domains (4.6%), and ten times the rate on .edu and .gov domains (both around 1%). That gap has widened since 2022, when the rates were much more uniform across domain types.
This makes sense. Content mills, affiliate sites, product description pages, and SEO-driven blogs operate on volume. AI cuts the cost of that volume dramatically. Academic and government sites face more scrutiny and have different incentives, so the pressure to automate is lower. The commercial web is simply more exposed to the economics that make AI writing attractive.
The linguistic tells that give it away
Detection models pick up on more than just vocabulary. Comparing 2023 to today, a few specific patterns have become significantly more common across the web:
- Em dash usage has roughly doubled
- Oxford comma frequency is up 63%
- AI-typical vocabulary words like ‘delve’, ‘interplay’, and ‘testament’ have more than doubled in usage
- Negative parallelism constructions have nearly tripled, though they remain relatively rare
None of these signals, taken alone, prove a document is AI-written. But at scale, across hundreds of thousands of pages, they form a consistent pattern that detection models can read with reasonable confidence.
Why this matters for developers and product teams
If you’re building anything that relies on crawled web data, whether it’s a training dataset, a RAG pipeline, a search index, or a content quality filter, this finding is directly relevant. A third of recently published pages carrying AI authorship signals means the feedback loop is already in motion. Models trained on web data are increasingly training on content generated by earlier models. The signal-to-noise problem for high-quality human-generated text is getting harder, not easier.
Tools like Open Pangram, Originality.ai, and GPTZero are trying to keep pace, but Pew’s own researchers note that no detection model is perfect. The more useful takeaway is systemic: the composition of the public web has changed, and products or research that treat web text as a proxy for human expression should account for that.
