Claude Opus 5.5 Cuts Em-Dash Use 99%, But Still Shows 2,548 AI Tells

What changed

Anthropic shipped Claude Opus 5.5 last week. An analysis by Graphite found that its use of em dashes fell 99% compared with its predecessor, while still identifying 2,548 other AI writing tells.

Why it matters

The change targets one of the easiest ways readers spot machine-written prose: a repeated punctuation habit. That makes Claude Opus 5.5 sound more natural at a glance, especially in routine writing where small stylistic clues can shape trust.

But the larger finding is less flattering. A model can remove one conspicuous habit and still leave thousands of detectable patterns behind. For teams using AI to draft customer messages, product copy or internal documents, “sounds human” is therefore a moving target, not a clean quality verdict.

Graphite’s analysis does not identify the predecessor, its methodology or sample size, so the 2,548 tells should be read as the result of this analysis rather than a universal score for every prompt. The practical takeaway is straightforward: Claude Opus 5.5 may be harder to identify from punctuation alone, but more natural wording does not mean the writing has become indistinguishable from a person’s.