Is Claude detectable?

Three different questions get asked as one. Whether Claude's watermark can be read, whether AI detectors flag Claude text, and whether anyone can prove you used it.

Detection · 3 min read

"Is Claude detectable" is three separate questions wearing one coat, and they have different answers. Sorting them out is more useful than any single yes or no.

1. Can Claude's watermark be read by anyone?
2. Do AI detectors flag Claude's writing?
3. Can someone prove you used Claude?

1. The watermark: not by anyone outside Anthropic

Claude embeds a statistical watermark in text from models launched on or after 2 August 2026. Anthropic has said it will publish detection details in forthcoming technical documentation. That documentation does not exist yet.

A statistical watermark is not a public checksum. Reading it requires the key, and the key is Anthropic's. So today:

• No website can check text for Claude's mark.
• No browser extension can.
• No university, publisher or plagiarism service can.
• We cannot.

This will change when Anthropic publishes. Right now the answer is a clean no.

2. Detectors: yes, but they are not reading the watermark

AI detectors flag Claude output frequently. They also flag GPT output, Gemini output, and a substantial amount of human writing.

They work on perplexity — how predictable each word is given the ones before it — and burstiness — how much sentence length varies. Model prose scores low on both: predictable word choices, uniform sentence rhythm. Detectors measure that and return a probability.

This is a style judgement, not a provenance reading. Which is why:

• It produces false positives on non-native English writers, whose vocabulary is often more regular.
• It produces false positives on technical and legal writing, which is deliberately uniform.
• Two detectors routinely disagree on the same passage.
• Editing model output until it reads naturally will lower the score, without touching any watermark.

So Claude text is "detectable" in the sense that a detector will often flag it. It is also detectable in the sense that a reader will often notice — delve, tapestry, Moreover at the start of three consecutive paragraphs, and em dashes everywhere.

3. Proof: weaker than people assume

Even setting aside that nobody can currently read the mark, consider what it would establish if they could.

Anthropic's own position is that a detected mark means content may have been processed by Claude. Not authored by. Processed by.

Text you wrote entirely yourself and pasted into Claude for a grammar check carries the same mark as text Claude drafted from nothing. The mark cannot distinguish those. Neither can a detector score.

That gap matters if you are on the receiving end of an accusation. The evidence available — now or after Anthropic publishes — does not separate "Claude wrote this" from "Claude edited this" from "this person writes in a regular style".

What is actually knowable today

The first two are the only rows where a tool can give you a number that means something. A scan returns exact counts per category — if it says four zero-width characters, there are four.

If you have been flagged

The useful moves are not tool-shaped.

Detector scores are probabilistic and contested, and institutions increasingly know it. Draft history, version snapshots and the ability to discuss your own argument in person are worth more than any score, in both directions. If you can show the work developing, that is evidence; a detector percentage is an estimate.

And if you have a disclosure obligation, no tool resolves it. Removing markers does not change what you did.

Run a free scan to see what is exactly countable in your text.

Related: how AI detectors work · why detectors flag human writing · Claude's watermark and Turnitin · how to check for the watermark