Most “how to spot AI writing” advice treats every model as one thing. It isn't. GPT, Gemini and Claude have noticeably different habits, and if you read a lot of each you start telling them apart the way you'd tell apart two colleagues' emails.
This is about Claude specifically — what its prose actually looks like, which tells hold up, and which ones people repeat that don't.
Short version
Claude's signature isn't a word, it's a posture: it hedges, it balances both sides, and it structures everything. Look for a piece that carefully refuses to commit and organises itself into tidy parts even when the question didn't call for it.
What Claude's writing actually sounds like
These are the patterns that hold up across a lot of samples. None of them is proof on its own — people write this way too, especially people who write carefully — but stacked together they're distinctive.
It hedges, constantly
“It's worth noting…”, “That said…”, “I should mention…”, “while X, it's also true that Y”. Claude is trained to avoid overclaiming, and that shows up as a steady drip of qualifiers. A human expert writing about their own field is usually more willing to just assert something. A piece that hedges every third sentence is either a very cautious writer or a model.
It gives you both sides whether you asked or not
Ask for an argument and you often get a balanced overview instead. There's a characteristic shape: the case for, the case against, then a synthesis that lands somewhere reasonable in the middle. Genuine opinion writing usually picks a side early and spends its length defending it.
It structures everything
Headers, numbered steps, bolded lead-ins on bullet points, a summary at the end. Claude reaches for structure by default, even for a question that would naturally be answered in one paragraph. If a 400-word answer arrives with three subheadings and a closing recap, that's a signal.
The lead-in bullet pattern
Specifically this shape: “Clarity: the writing should be easy to follow.” A bolded one-or-two-word label, a colon, then an explanation. It's a genuinely good format, which is why it's everywhere in Claude output and comparatively rarer in human drafts, where bullets tend to be scrappier and less parallel.
It restates the question first
A short opening that reframes what was asked before answering it. Humans usually just start answering, because they already know what they asked.
The tells that don't work for Claude
Some of the standard advice actively misleads here.
Em dashes
Claude does use em dashes, and for a while that was treated as a giveaway. It stopped being reliable for two reasons: models got tuned away from it once it became a meme, and plenty of good human writers use them heavily. Judging by punctuation alone will flag a lot of innocent people.
“Delve” and friends
Word-level tells decay with every model release. The vocabulary that felt distinctive in 2024 is largely gone from current output. If your detection method is a banned-words list, it has a shelf life measured in months.
Politeness
Claude's warmth in a chat window is a product of the assistant format, not of the prose it produces. Ask it for a technical document and you get a technical document. Judging a finished essay by whether it sounds friendly tells you nothing.
Why Claude is often harder to catch than GPT
In practice, Claude-written prose tends to trip detectors slightly less often than some other models, and the reason is mundane: its sentences vary more in length. Most statistical detectors lean on perplexity and burstiness — roughly, how predictable each word is and how much sentence length varies. Output that mixes a 30-word sentence with a 6-word one looks more human on those measures than output that marches along at a steady 18 words.
That doesn't make it undetectable. It means the signal is more often in the structure and the hedging than in the raw statistics, which is why reading the piece still matters.
Checking it instead of guessing
If you'd rather measure than squint, our Claude AI detector scores text against Claude's specific patterns and shows you the sentences that pushed the number up. The reasoning is the useful part — a bare percentage tells you nothing you can act on or argue with.
Before you accuse anyone
This matters more than any tell in this article. AI detection is probabilistic, and the people most likely to be wrongly flagged are the ones who write in clean, careful, well-organised prose: non-native English speakers who have been taught a formal structure, students who follow the essay template they were given, and anyone who edits heavily.
A high score is a reason to have a conversation, not evidence of dishonesty. Ask about the drafting process. Ask what they meant by a particular paragraph. That will tell you far more than a number will.
The honest summary
You can learn to recognise Claude's voice, and the tells above are real. But they describe a style, and styles can be imitated in both directions — a person can write like Claude, and Claude can be prompted not to. Treat the pattern as evidence to weigh, not a verdict to deliver.
