📰 Key Takeaways

In articles generated by OpenAI’s Astra model, phrases like “another dimension” show up at an unusually high rate, and the model favors hedging language like “may provide” or “can provide” when describing benefits. Its most telling signature is the “corrective frame” — defining an issue as “not just X” or “rather than relying on X, instead…” — a pattern that shows up over 100 times more often in Astra’s output than in human writing.

Interestingly, it looks like the major frontier models have all adjusted in response to criticism about overusing em-dashes: Opus 5.5’s em-dash usage dropped 99% compared to Opus 5, Astra uses 88% fewer than human samples, and Gemini 3.1 Pro has nearly stopped using them altogether. But Graphite points out that even though individual “tells” shift from version to version, the overall count of tells hasn’t actually gone down. In Druck’s words: “These tells aren’t disappearing — they’re being managed, shifted around.”

💬 JudyAI Lab Take

This story exposes the “writing fingerprint” of AI-generated text — even when a model deliberately fixes one tell (like slashing em-dash usage), the total amount of detectable signal doesn’t actually drop. It just moves from one feature to another, like the “corrective frame” sentence pattern, which shows up at 100x the rate of human writing.

That’s a signal worth sitting with if you’re building AI tools: trying to make AI text undetectable by stamping out one feature at a time might just be a losing game, because a model’s language habits are systemic — patch one tell and another pops up somewhere else. It points to a deeper design question too: instead of pouring effort into “making AI write more like a human,” maybe it’s worth asking what AI-generated content should actually be for, rather than just imitating human writing. For anyone building content generation tools or evaluating AI text quality, this kind of “style fingerprint” research gives you a real direction for detection and self-correction.

Next time you’re looking at AI-generated text, it’s worth checking for more than just em-dashes — watch for that “not just X” corrective framing too.

📅 Source Info

🔗 Further Reading