The Em Dash Retired. Nine New Tells Took Its Place.

For a year the em dash has been treated online as a smoking gun for AI writing, so much that clean, well-punctuated human prose started drawing suspicion too. This month an AI newsletter writer with a large following posted an updated file of tells meant to survive now that models have been tuned to avoid the dash itself. Ruben Hassid writes about artificial intelligence for 885,000 newsletter subscribers and roughly 900,000 LinkedIn followers, and in mid-August 2026 he posted on X a file he keeps updated to spot AI-generated text on sight. The list spread fast through tech blogs, and we added it to our trend board this week.

A separate look at EQ-Bench's public slop data put real numbers on that gap. The benchmark counts overused AI phrases per 1,000 words, sets the human baseline at 6.90, finds most models scoring between 10 and 40, and notes that only one specially tuned model has beaten that human baseline so far. That is the same kind of drift our own draft checker looks for in an ordinary paragraph, not a single flashy word but a steady hum of overused phrasing.

Wikipedia's own guide to spotting AI writing added a specific new pattern this month too. It notes that newer models often reach for indirect constructions, describing two things as associated rather than stating the relationship plainly with a word like of, for, or by, especially when paired with generic promotional vocabulary. On its own the habit proves nothing; stacked with a rule-of-three list or an unnamed source, it adds up. Run a paragraph through our fix-your-ai guide before you second-guess your own voice.

None of this changes the basic advice. A single tell is weak evidence, but density and repetition are not. GPTZero's own vocabulary page was refreshed this month with no standout new figures, and our phrase and dash charts are unchanged since last week. What is new is where to look. Check your drafts against the widening list, not just the em dash you were already told to delete.