Flagged as AI when you wrote it yourself
If you used Claude to proofread, translate or reformat your own writing, your text can carry Claude's watermark even though you wrote every idea in it. This is not a malfunction and not a rumour — it is stated by Anthropic, which says a detected mark means content may have been processed by Claude, is not fully conclusive, and does not establish that Claude was the original author[1].
That sentence is the single most useful thing on this site if you have been accused of something. It is Anthropic's own documentation, not an argument made on your behalf.
How you can be marked without cheating
- You wrote a draft and asked Claude to fix the grammar.
- You wrote in your first language and had Claude translate it.
- You asked Claude to shorten your own paragraph, or fix the formatting.
- You pasted your notes in and asked for a summary you then rewrote.
All of these are ordinary uses. Anthropic names proofreading, translating, summarising and converting files specifically as the reason a mark cannot be read as authorship[1].
What to say if you are challenged
Four things are true and checkable, and you can point to the source for each.
- A mark is not authorship. Anthropic states a detected mark is not fully conclusive and does not establish full provenance, because the tool is commonly used on text the person wrote[1]. Its technical explainer puts the limit in one sentence: a watermark cannot distinguish "Claude wrote this" from "Claude heavily edited this"[10].
- Ask what produced the evidence, and whether they could have it. Since 1 September 2026 a Claude watermark detection API does exist, in private preview[10]. Anthropic offers it to organisations that EU law obliges to check — regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups — and to enterprises under the same obligation, by application[1][12]. It is not on sale and it is not on the open web. So the question to ask is simply which tool produced the result. If the answer is a website, a plagiarism suite or an AI-detector score, it is not a watermark result. See the audit.
- Proofreading barely marks anything. Anthropic says the watermark only applies to words Claude chooses, so when Claude edits your writing there is very little for it to attach to — on a light grammar pass the mark can live only in the handful of corrections, which may be too few to register[10].
- An "AI detector" is not a watermark detector. The classifiers used by most schools guess from style, and have no access to any watermark. They are a different technology with different failure modes. The difference, plainly.
Keep your drafts. Version history in Google Docs or Word, or a folder of dated files, is the evidence that actually settles these disputes. It shows the work forming over time, which no watermark and no classifier can speak to.
Anthropic's list of who may apply for the detection API names educational organisations[1]. So a school or university is not, in principle, excluded from holding one, and we would be misleading you if we said otherwise. What we can say is what the list is for: it is scoped to bodies with an obligation under EU law to verify marking, access is granted by application, and Anthropic has published no list of who holds it. Ask, rather than assume — in either direction.
What we cannot help with
We cannot tell you whether a specific piece of text is marked. The only detector for Claude's text watermark is Anthropic's, it is in private preview, and it is not open to individuals[1]. Our scanner checks for hidden characters, which is a different and much narrower thing. Anyone offering you more than that right now is selling something.
Sources
Last verified against primary sources: