Graphite Finds 13,000 AI Writing Tells in Claude Opus 5.5 and OpenAI Astra
Updated
Updated · TechCrunch · Oct 1
Graphite Finds 13,000 AI Writing Tells in Claude Opus 5.5 and OpenAI Astra
1 articles · Updated · TechCrunch · Oct 1
Summary
Graphite identified 13,000 phrases that appeared at least twice as often in AI-generated prose as in human writing, suggesting frontier models still leave broad linguistic fingerprints.
The study compared 10,000 pre-ChatGPT human articles with AI rewrites generated from summaries, letting researchers isolate model-specific word choices and sentence patterns.
Claude Opus 5.5’s strongest tells included “dependable” at 23 times human frequency and “this matters” at 116 times, while Astra overused corrective framing such as “not simply X” more than 100 times as often.
Labs have curbed older giveaways like em-dashes—Opus 5.5 used them 99% less than Opus 5, and Astra 88% less than human samples—but Graphite said new tells keep replacing the old ones.
Graphite said Claude models are moving closer to human word distributions over time, while GPT models are drifting further away, underscoring how hard it remains to make LLM prose fully humanlike.