The tool
What this Claude detector actually does
It reads the text you paste and writes back what is observable in it: hedged phrasing, uniform sentence construction, structure that repeats, openings that qualify before they say anything. It does not tell you who or what wrote the text, and it does not produce a percentage.
That is a deliberate limit, not a missing feature. The analysis runs on our server, so your text is sent to us, with 5 free runs every rolling 24 hours and an input of 1 to 12,000 characters. What comes back is a written reading you can check line by line against the page in front of you.
The honest part
No detector can tell you Claude wrote something
This is not our opinion. It is what the competing detectors say about themselves, on the pages selling the detection. AIDetego runs a Claude detector page and states on it that "Exact attribution to a specific model (Claude vs. ChatGPT) is currently not possible", adding that "no tool can reliably do this".
Plagly.ai lists five Claude tells and then writes: "None of these is proof on its own, and careful human writers legitimately share several of them." Detector-checker.ai says its tool "does not identify the source model" and cannot distinguish Claude Sonnet, Opus and Haiku from each other. Once you have read those three sentences, a page promising to name the model is selling something its own industry says does not exist.
How it works
How to check text in three steps
Use it on your own writing. That is where a reading of the signals is worth something.
- Step 1
- Paste a passage. There is no minimum length here, so a paragraph works and is often easier to act on than a whole document.
- Step 2
- Add an instruction of up to 300 characters if you want the reading pointed somewhere, then press Analyze.
- Step 3
- Check each finding against the text yourself. Anything that survives that check is real; anything that does not cost you ten seconds.
What to look for
The patterns detectors associate with Claude text
Two of the detector pages publish their signal lists, and the lists overlap. As editing prompts they are useful, whoever wrote the draft.
- Hedging
- Plagly names "it is worth noting", "generally", "in many cases", "I would suggest". AIDetego names "it seems", "potentially", "tends to". Qualifiers in front of every claim.
- Qualified openers
- Sentences that begin with a clause softening what follows, so the reader reaches the point late or not at all.
- Even structure
- Numbered steps, headers and neat summaries in places that did not need them, and a thesis-then-counterargument-then-synthesis shape repeated section after section.
- Uniform quality
- AIDetego calls it "Consistently high readability throughout". Nothing rough, nothing sharp, no sentence noticeably better or worse than its neighbours.
Why no number
Why this Claude detector gives no percentage
A percentage looks like a measurement and is treated like evidence. It gets pasted into an email to a student or a writer, and at that point a statistical guess about sentence regularity has become an accusation.
The signals above are also the signals of careful human writing, which plagly.ai concedes in its own words. A person who hedges, structures and edits well will read exactly like the thing being detected. A written finding invites you to look; a number invites you to conclude.
The watermark question
Does this detect a Claude watermark?
No. This route detects no watermark of any kind, and it is worth being precise about why the question exists. A page at overchat.ai asserts that Claude models add "an invisible statistical pattern" to their output by "subtly influencing token selection during generation". We cannot verify a claim about another company's models and we neither confirm nor deny it.
What that same page then admits is the part worth keeping. Its own scanner looks for zero-width characters, hidden Unicode, punctuation style and formatting residue, and it says: "Finding one of these artifacts therefore does not prove that text was written by Claude, and removing them does not necessarily remove Claude's watermark." Artifacts are not authorship, and a statistical pattern inside a model is not something any text box can see.
Who it is for
Who this is for, and who it is not
It is for writers checking their own drafts, editors deciding whether a piece needs another pass, and anyone who wants the tells named so they can be edited out. In that use the findings are immediately testable, because you have the draft.
It is not for deciding whether someone else cheated. No output here supports that decision, and we would rather say so than let a page imply otherwise. If that is the question you have, it is answered by talking to the person and looking at their drafts, not by a tool.
Limits, stated
What this Claude detector will not do
No verdict, no percentage, no confidence score, no bands, no sentence-level highlighting. No claim that Claude, ChatGPT or any other model wrote your text. No accuracy figure, because nothing has been measured here and repeating someone else's benchmark would describe a different tool.
No watermark detection, no hidden character scan on this route, no file upload, no unlimited scans, no API, no account and nothing stored. Competing pages advertise most of that list, and several of them also publish the sentence explaining why the headline claim cannot hold.