Pew estimates a third of recent web pages carry AI text, as Pangram points to guardrails
A Pew Research study found AI-writing signals in 10 percent of 10,000 web pages. Detection company Pangram argued that post-training safety rules make model text easier to spot.
Artificial Intelligence··Night
Pew's detection findings
A new Pew Research study suggests that a significant portion of recent internet content is machine-generated, with over a third of English-language web pages published since ChatGPT's release showing signs of AI writing. The researchers analyzed a July 2026 sample of 10,000 pages from Common Crawl using Open Pangram detection tools. While the overall sample showed clear AI signals in about 10 percent of pages, Pew noted that this figure includes older content, and the share rises sharply for newer pages.[1]
The role of model guardrails
The detection method behind the Pew findings relies on the structural predictability of commercial language models. Pangram chief technology officer Bradley Emi argued in a company blog post that safety and behavior training shrinks the expressive range of models like ChatGPT, Claude, and Gemini. According to Emi, this post-training process produces a recognizable mode-collapse signature that detectors can spot, whereas base models without guardrails, narrow fine-tunes, and broken outputs write with more variety and often slip past the checker.[2]
Guardrails shape the visible web
These findings indicate that the current visibility of AI writing on the web is closely tied to the specific safety guardrails imposed by major vendors. Because the Pew study relies on Pangram's detector, the estimated share of machine-generated pages primarily reflects the recognizable output of post-trained commercial models. If developers shift away from these restrictive behavioral guardrails, or if users increasingly deploy unaligned base models, the web may contain even more AI text that evades this specific detection method.[1], [2]
