Study finds common phrases that can reveal AI written text

In short: A new analysis says today’s AI writing still has repeating habits that can make it easier to spot.

What's going on

A study by the marketing firm Graphite looked for “tells” in AI generated writing. A “tell” is a small clue that gives something away, like a repeated catchphrase.

Graphite compared human written articles with versions rewritten by several well known AI models from summaries. It used 10,000 articles published before ChatGPT as the human baseline, then measured which words and phrases showed up much more often in the AI versions.

The researchers said they found about 13,000 phrases that appeared at least twice as often in AI text as in human text. One example in the report is Anthropic’s Claude Opus 5.5, which used the word “dependable” 23 times more often than the human samples. The model also used “this matters” 116 times more often than people did, and “why X matters” 92 times more often.

OpenAI’s Astra showed different habits. Graphite said Astra often used phrases like “another dimension,” and it frequently softened claims with wording like “may provide” or “can provide.” The study also found Astra relied heavily on “corrective framing,” such as “not simply X” or “rather than relying on X,” which Graphite said showed up more than 100 times as often as in human writing.

What to watch

Graphite’s Greg Druck told TechCrunch that AI makers can remove famous giveaways, like overusing em dashes, but new patterns keep appearing in each model update. For readers, editors, and teachers, the key question is whether these clues stay stable enough to help flag AI text, or whether they shift too quickly to be reliable.

Source: TechCrunch AI