Learn · Claude
Does Claude
watermark its text?
Updated 24 August 2026
Yes — future Claude models will carry a SynthID-style word-choice watermark. What was announced, when it applies, and what detection can and cannot show.
The short answer
Yes — for future models. On August 14, 2026, Anthropic announced that future Claude models will generate text carrying a statistical watermark, describing the system as a version of Google DeepMind’s SynthID-Text. Anthropic linked the change to the EU AI Act’s transparency obligations for AI-generated content.
Three details matter before anything else: the watermark is carried by patterns in Claude’s word choices, not by hidden characters or metadata; it contains no information about who used Claude; and checking for it requires Anthropic’s keyed detector — nobody can spot it by reading.
What the watermark actually is
When a language model writes, it repeatedly chooses the next word from several plausible options. SynthID-style watermarking nudges those choices using a secret key, so that across a long passage the selections form a pattern the key holder can later test for. Nothing is added to the text — the sentence reads exactly as it would otherwise.
Because the signal accumulates across many small choices, it is strongest in long, varied prose and weakest in short or tightly constrained output such as exact factual answers, maths, or code, where the model has little genuine choice. For the full mechanism — tournament sampling, keyed scoring, and detection — see our technical explainer.
When it applies — and when it does not
The announcement covers future Claude models, so text generated before the change is not described as marked. Where Claude chooses the output words — including translation — the output carries the watermark. Light human editing probably does not remove it completely, according to Anthropic, while replacing every word does; the SynthID-Text paper likewise reports that edits and paraphrasing weaken generative watermarks.
Anthropic announced a detection API alongside the watermark but described it as forthcoming. Until it ships, there is no public way to test a passage for Claude’s watermark — and no third-party tool can honestly claim to detect it, because detection requires the secret key.
What a positive detection result would not mean
A watermark detector is a provider-specific statistical test, not an authorship oracle. A high score supports the conclusion that Claude was likely involved somewhere in producing the passage. It cannot establish which person or account used Claude, whether unmarked text was written by a human or a different model, or who owns and is responsible for the final work.
That distinction matters in practice: a document drafted by a person and later polished by Claude, and a document generated wholesale, can both carry signal. The test speaks to involvement, not to how the text came to be.
Quick answers
Does Claude hide invisible characters in its text?
No. Anthropic states the watermark is carried by patterns in word choices, not by hidden characters, metadata, or extra tokens. Invisible Unicode characters in a draft come from other sources, such as copy-paste artefacts or older marking tools.
Do older Claude conversations carry the watermark?
Anthropic's announcement describes the watermark as a feature of future Claude models. Text generated before the change is not described as watermarked.
Can the watermark identify who used Claude?
No. According to the announcement, the mark carries no user, organisation, or chat identity. Detection can only estimate whether Claude was involved in producing a passage.
Does copying and pasting remove the watermark?
No. Because the signal lives in the words themselves, it survives copy-paste, format changes, and file conversions. Only changing the words changes the signal.
Working on a draft now? The workspace combines character cleanup, semantic rewrite, and humanize. It is free to try without an account, and it does not promise a particular detector outcome.
Open the text workspace →Sources
- Anthropic announcement — How Claude’s text watermark works · August 14, 2026
- Nature paper — Scalable watermarking for identifying large language model outputs · October 24, 2024
- EU AI Act — Regulation (EU) 2024/1689 · Article 50 transparency obligations