Full Breakdown
New Research Maps Thousands of “Tells” in AI-Generated Writing
By Drooid · · How we work
Study Overview
A marketing firm, Graphite, released a large-scale analysis of the linguistic habits of frontier large language models (LLMs). Researchers compared 10,000 pre-ChatGPT articles (human-written) with AI-rewritten versions produced by several leading models. The goal was to identify words and constructions that appear disproportionately in machine-generated text, which Graphite calls “tells.”
Key Findings
- The analysis uncovered roughly 13,000 phrases that occur at least twice as often in AI output as in human writing.
- The phrase “this matters” is the most frequent tell, appearing 116 × more often in Claude Opus 5.5 text than in the human baseline.
- OpenAI’s Astra model favors “may provide”/“can provide” hedges and the phrase “does not establish,” which is 275 × more common than in Claude outputs.
- All surveyed models have dramatically reduced em-dash usage: Opus 5.5’s em-dash frequency is 99 % lower than its predecessor, and Astra’s is 88 % lower than human samples.
Data Highlights
- 13,000 identified tells across models.
- 116 × frequency for “this matters” (Claude Opus 5.5).
- 23 × frequency for “dependable” (Claude Opus 5.5).
- 92 × frequency for “why X matters” (Claude Opus 5.5).
- 275 × frequency for “does not establish” (OpenAI Astra).
Official Statements & Responses
Greg Druck, Graphite’s chief AI officer, emphasized that while models are shedding well-known tells, new ones continuously emerge. He noted, “It’s not like the tells are decreasing,” “It’s not like the tells are decreasing,” — Greg Druck and added, “A general hypothesis I have is that the labs are less able to control some of these things than you might expect,” “A general hypothesis I have is that the labs are less able to control some of these things than you might expect,” — Greg Druck.
Implications
The persistence of model-specific linguistic fingerprints suggests that AI-detector tools will remain relevant, even as developers refine output style. The study also highlights a “whack-a-mole” dynamic: eliminating one tell often leads to the emergence of another, underscoring the difficulty of fully humanizing LLM prose.