# New research: Avoid these 3-word phrases if you don’t want to sound like AI > Source: > Published: 2026-10-07 10:00:00+00:00 With each new model, [artificial intelligence](https://www.fastcompany.com/section/artificial-intelligence) is getting closer and closer to replicating human writing—but it’s not there yet. Anthropic says its newest model for Claude, [Opus 5.5](https://www.anthropic.com/claude-opus-5-5), “communicates more naturally than prior models.” But that revamped communication style comes with its own set of hallmarks, and a new study lays out the writing trends in today’s most advanced AI models. [Researchers at Graphite](https://graphite.io/five-percent/research/ai-tells-opus-5-5-update), a research-based growth agency, compared 10,000 articles generated by Opus 5.5 to 10,000 human-written articles on the same topics. This generated a list of tells, which they define as words, phrases, and patterns that occur at least twice as often in AI writing as in human writing. Opus 5.5 has fewer tells than previous Opus models, but it still has upwards of 2,500. If you want to keep your writing from sounding AI generated, here are the words and phrases to avoid, according to the study. Though single words aren’t likely to get your prose pegged as generated by AI, there are still some buzzwords that scream Opus 5.5. Here are the top five culprits, along with their frequency compared with usage in human writing. When it comes to two-word phrases, Opus 5.5’s common function of breaking down concepts becomes especially clear. The top tells in this category appear infinitely more times in AI writing than in human-written content, meaning that while Opus 5.5 used these phrases hundreds of times, humans didn’t employ them even once. Beyond specific words and phrases, the researchers identified style choices that Opus 5.5 employed more often than humans. The most pronounced among them included significantly greater lexical diversity, longer words, formulaic closes, higher word counts, and increased use of commas. Meanwhile, the most stereotypical tell of AI writing—overuse of em dashes—has all but disappeared from current models. Opus 5.5 uses just 0.015 em dashes for every 1,000 words, a 99% drop from Opus 5. That’s less than 0.01 times the rate that humans use em dashes. This means that nowadays, they’re more likely a sign of human-written content than of AI-generated writing. Opus 5.5 also tones down previous models’ use of mannered prose, which is flowery language and metaphor, in place of direct statements. Using their own scale, researchers determined that Opus 5.5 uses mannered prose 37% less than Opus 5. But the model still uses mannered prose far more than humans, opting for flourish over direct language 1.6 times as often as human writers do.