AI writing tells persist despite lab efforts

AI writing tells persist despite lab efforts
A study by marketing firm Graphite found 13,000 phrases at least twice as common in AI-generated content versus human writing. Researchers had frontier models rewrite 10,000 pre-ChatGPT articles from summaries, then compared word and phrase frequencies. Claude Opus 5.5's biggest tell is using "this matters" 116 times more often than humans, while OpenAI's Astra overuses "corrective framing" constructions like "not simply X" more than 100 times the human rate. Labs have largely eliminated em-dash overuse, but Graphite's chief AI officer says new tells keep emerging with each model version, and overall tell counts remain steady.
Read the original article →