Anthropic Adds Text Watermarking to Claude to Comply With EU AI Act
Anthropic says future Claude models will embed an undetectable watermark in generated text to comply with the EU AI Act.
Quick answer
How does Anthropic's new watermarking system for Claude's text output work?
Anthropic says future Claude models will embed a watermark in generated text by altering how the model selects among equally likely words, using a method called SynthID-Text. The company says the watermark is undetectable to readers, adds no cost or delay, and carries no information that could identify a user, organization, or chat.
Key takeaways
- Anthropic says future Claude models will generate text containing a watermark to comply with the EU AI Act, which as of August 2, 2026 requires AI providers serving the EU market to mark AI-generated content.
- The company says the method, a version of Google DeepMind's SynthID-Text technique, works by changing the source of randomness Claude uses when picking among equally suitable words, rather than adding hidden characters or extra tokens.
- Anthropic says the watermark has no effect on output quality, speed, or cost, and carries no information that could be traced to a specific user, organization, or conversation.
- Anthropic says it is applying watermarking globally at launch because it does not yet have a durable way to scope it by region, and plans to extend it to older Claude models over the coming months.
- Anthropic says it is releasing a detection API in private preview, initially for regulators, law enforcement, media, fact-checkers, researchers, and similarly obligated enterprises, with plans to expand access over time.
Anthropic says future Claude models will generate text carrying a watermark, a change the company says it is making to comply with the European Union's AI Act. In a post explaining the mechanism, Anthropic says it is one of several major AI providers implementing the requirement.
How the watermark works
Anthropic says large language models generate text one word at a time, frequently facing choices between multiple words that are equally sensible, such as picking "overcast" or "grey" to describe the weather. In those low-stakes moments, the company says, an unwatermarked model settles the choice with an arbitrary random number. Watermarking instead uses a cryptographic key together with the preceding words to determine the choice. Anthropic says this leaves a statistical pattern across a piece of text that is undetectable to a reader but can be checked by anyone holding the key, allowing them to estimate the probability that Claude produced the text.
Anthropic says the method is a version of the SynthID-Text approach published by Google DeepMind in a 2024 Nature paper, building on a design principle first proposed by Scott Aaronson in 2022.
Where it applies less
According to Anthropic, the watermark has less to act on in text where word choice is constrained. The company cites factual statements, such as completing "Isaac Newton's most famous work was called Principia," and code, where an incorrect token can break a program, as places where watermarking is sparser. Anthropic says detection also becomes less reliable on short passages, since fewer word choices provide less information, and that light proofreading of human-written text may leave too little Claude-generated content to register. The company says translations are fully watermarked, since Claude chooses every word.
Anthropic says heavy editing or a complete rewrite can remove the watermark, while light edits generally will not.
Rollout and detection
Anthropic says it signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, alongside roughly 190 total signatories. The company says it is applying the watermark globally at launch because it does not yet have a durable way to limit it by region, and that Claude models released before August 2, 2026 will receive watermarking over the coming months under the law's transition period.
The company says it is releasing a detection API in private preview, currently available to organizations it describes as eligible under EU law, including regulators, law enforcement, media, fact-checkers, researchers, educational organizations, and EU civil society groups, as well as enterprises with their own compliance obligations. Anthropic says it plans to expand access over time.
For image and other file types such as .png, .jpg, and .svg, Anthropic says Claude attaches a cryptographically signed content credential in the file's metadata using the C2PA industry standard, rather than a watermark, noting that this metadata states Claude was involved without altering or hiding anything in the file itself. The company says the watermark does not determine ownership of an output or change a user's rights under its terms.
Source: Anthropic, "How Claude's text watermark works," published Aug. 14, 2026, updated Sep. 1, 2026.
Frequently asked questions
- What is Claude's text watermark?
- Anthropic says it is a pattern created by altering the randomness Claude uses when choosing between equally likely words during generation, detectable with the company's key but not by ordinary readers.
- Why is Anthropic watermarking Claude's text?
- Anthropic says it is complying with the EU AI Act, having signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026 along with other major AI providers.
- Does watermarking slow down or raise the cost of using Claude?
- No. Anthropic says watermarking has a negligible effect on speed and produces no extra tokens, so it does not change the cost of serving or using the model.
- Can the watermark identify who wrote a piece of text?
- No. Anthropic says the watermark contains no information that could identify a user, their organization, or their chats with Claude.