Marketing firm Graphite discovered that frontier AI models like Claude and GPT still rely on 13,000 distinct writing tells and phrases that appear at least twice as often in AI content as in human content. Researchers analyzed 10,000 pre-ChatGPT articles and compared them against AI-rewritten versions to identify unique quirks, such as Claude Opus 5.5 frequently using the phrase "this matters" and OpenAI's Astra employing corrective framing. While labs have successfully stamped out earlier giveaways like em-dashes, new verbal habits continually emerge across model updates.
Graphite chief AI officer Greg Druck noted that Claude models are actually getting closer to human word distribution over time, whereas GPT models are moving further away. The study found specific linguistic markers for different systems, with Opus 5.5 favoring the word dependable and Astra heavily relying on hedges like "may provide." Druck suggests that AI labs have limited ability to completely control these patterns because billions of parameters make full stylistic oversight difficult.
This research matters because it exposes the persistent linguistic footprints that distinguish machine-generated prose from human writing despite ongoing updates from major AI labs.