Quick answer
Since summer 2026, text generated by new Claude models carries an invisible watermark. Anthropic explained it in an article published on August 14, 2026.
The reason is regulatory. In July 2026, Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content, tied to the AI Act. The code requires marking the outputs of models released after August 2, 2026. Other major providers signed it too and will implement their own watermarks.
Fable 5.1 and Opus 5.5 are covered.
How it works, without the math
A language model writes one word at a time. At each word, it has several candidates. Anthropic uses the sentence "The weather today was cold and…". The next word is very unlikely to be "sugary", but "overcast" and "grey" are both quite likely. For the reader, the choice between them changes nothing. Normally, that choice is settled by a random number.
The watermark doesn't change the fact that the choice is random. It changes the source of the randomness. Instead of an arbitrary random number generator, the model uses a secret key and the few words that came before to settle the pick. Someone holding the key can then check whether the sequence of words is consistent with those choices and assign a probability that Claude wrote the text.
Anthropic uses a Monopoly analogy that boils down to this: instead of rolling dice, players read the digits of pi from a given position. The game stays just as random for the players. But someone who knows pi can look at the moves played and tell whether the game probably used pi.
The method is a version of the SynthID-Text approach published by Google DeepMind in 2024.
What the watermark isn't
Anthropic's article answers several concerns one by one:
- No hidden characters. Nothing is added to the text.
- No quality loss. Anthropic says it saw no impact internally. Google DeepMind, which introduced the technique, found no statistically significant difference in Gemini user ratings.
- No extra cost. No extra tokens, same price, negligible impact on speed.
- No identification. The watermark contains nothing about the user, their organization or their chats.
- No change in ownership. The watermark says nothing about who owns the text or who's legally responsible.
What about code?
If you use Claude Code, the natural question is: is my code marked? The answer is "very little".
The watermark only applies where several choices are equally good. When only one word is right, such as the answer to "2 + 2 =", nothing changes. And code very often has to be exact. Anthropic therefore writes that code generally carries less watermarking than other text, and that the effect on the code produced is negligible. Comments in code, where word choice is free, can carry it.
Special cases
Proofreading. If you ask Claude to fix only the grammar of text you wrote, the watermark can only live in the few corrections. There's a good chance it's undetectable. The more Claude writes, the more room there is for the watermark.
Translations. They carry a watermark, since every word is chosen by Claude.
Short, factual text. Detection works poorly on small samples and on very factual passages, where there are few free choices.
Rewriting. Light editing probably won't remove the watermark. A complete rewrite where every word is replaced will. Anthropic notes that at that point, it's debatable whether the text is still "AI-generated".
Images and files. For files like .png, .jpg or .svg, it's not a watermark but a signed note in the metadata, using the open C2PA standard, saying the file was made or processed with Claude.
Who can check?
Anthropic is opening a detection API in private preview. For now it's limited to organizations eligible under EU law: regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups. Enterprises that must verify watermarking for their own AI Act compliance also have access. Access should expand over time.
In other words, an individual or a teacher can't currently check a text themselves with the official tool.
Why it's global
The watermark is applied everywhere, not just in the EU. Anthropic explains it doesn't yet have a durable way to scope it by region. Models released before August 2, 2026 get a transition period, and the watermark will be added to them over the coming months.
The takeaway
For a developer, the change is nearly invisible: no cost, no quality loss, a negligible effect on code. For someone publishing text written with Claude, such as an article or documentation, that text is now potentially detectable by some parties. It's not a reason to worry, it's a reason to be open about how you use AI.
Next steps
- Claude Opus 5.5: what actually changes: the first Opus launched with the watermark
- Fable 5.1 and Mythos 5.1: the announcement that also details AI Act compliance
- Security best practices: what to know before handing data to a model
- 8 Claude Code myths: separating fact from fiction