Opus 5.5 has a new set of AI writing tells

AI models finally stopped overusing em-dashes, but they still can't stop telling you why things matter.

Marketing firm Graphite tested top AI models against 10,000 pre-ChatGPT articles to identify machine writing quirks, according to TechCrunch. Researchers found 13,000 phrases at least twice as common in AI content as human writing. Whenever AI labs patch old giveaways, new tics pop up to replace them.

Why it matters: Claude models are slowly moving closer to natural human word choices, but OpenAI's GPT models are drifting further away. AI labs boast about natural tone in releases like Opus 5.5, Sol, and Luna. Yet with billions of parameters to manage, researchers say labs simply can't catch every pattern that slips through.

Know this: Here are the easiest habits to spot right now:

• Claude Opus 5.5: Drops the word "dependable" 23 times more often than human writers, uses "this matters" 116 times more often, and loves saying something "is more than an X, it's a Y."
• OpenAI's Astra: Loves "another dimension," hedges claims with "may provide," and uses "not simply X" framing over 100 times more often than humans.
• Gemini 3.1 Pro: Almost completely wiped em-dashes from its writing, while Opus 5.5 cut them by 99% compared to Opus 5.

The next time a piece of text insists "this matters," you'll know who wrote it.