New research: Avoid these 3-word phrases if you don't want to sound like AI

A short news item by Jude Cramer for Fast Company. It reports a study from Graphite, a growth agency, that compared 10,000 articles written by Claude Opus 5.5 with 10,000 written by people on the same topics. Graphite calls a word or phrase a “tell” when the model uses it at least twice as often as the human writers did. Opus 5.5 has fewer tells than earlier Opus models, and still more than 2,500.

The article lists the worst offenders. Single words: “dependable” (23 times the human rate), “clearer”, “matters”, “quietly” and “practical”. Pairs: “matters because” (146 times), “especially valuable”, “this matters”, “article explains” and “builds trust”. Then five three-word phrases the model used hundreds of times and the human writers never did: “helps to understand”, “article explains what”, “this matters because”, “taken together these” and “this guide walks”. Em dashes are nearly gone (0.015 per 1,000 words, down 99% from Opus 5), so the article says an em dash is now more likely the mark of a person. Flowery, metaphor-heavy prose is down 37% from Opus 5 and still 1.6 times the human level.

Graphite’s own write-up (“AI Tells: Opus 5.5 Update”, 1 October 2026) backs the headline numbers: 2,548 tells against 2,666 for Opus 5, the em-dash rate, the scores for mannered prose, and “dependable” and “this matters” at the multiples quoted. Most of the other words, and all five three-word phrases, do not appear in the text of either Graphite page. They may come from its data file or its interactive charts, which were not checked for this page.

The article also leaves out the study’s own limits. The human articles all date from before ChatGPT launched in November 2022, so ordinary drift in web writing since then can show up as a tell (Graphite says so itself). Every model article was written from a summary of the human one, under a single prompt that asked for a professional yet friendly tone; Graphite expects a different prompt would change the tells. The score for mannered prose was given by another model, Claude Opus 5, on a sample of 1,000 topics. Graphite’s authors say individual tells should not be used to judge who wrote something, which is what the headline invites. It is a company research post (Graphite’s site offers SEO and AI-search services) and has not been peer reviewed.