Anthropic is adding invisible watermarks to text generated by supported Claude models as part of its commitments under the EU AI Act. The company has now explained how the system works, what happens after editing, and why AI-generated code will carry a much weaker watermark.
Anthropic is shedding more light on a new system designed to identify text produced by Claude, after the company’s decision to introduce AI watermarking sparked debate among users.
The company says the watermark will not appear as a visible mark or change the way a response reads. Instead, Claude will make subtle choices while generating text that create a machine-readable pattern. People reading the response should not be able to tell whether it has been watermarked, while someone with the appropriate detection key can check for the embedded signal.
Anthropic has now outlined the technology and its limitations in a blog post, addressing questions around editing, proofreading and code. The move follows its announcement that Claude would adopt watermarking to meet requirements under the European Union’s AI Act Transparency Code.
How Claude’s watermarking works
The system will use the SynthID-Text technique developed by Google DeepMind and described by the research team in 2024. Anthropic says it plans to make a watermark detection API available as well.
The basic principle relies on choices that do not materially affect the meaning of an answer. For instance, when several words could work equally well in a sentence, Claude can use those choices to create a statistical pattern. That pattern remains invisible to the reader but can be identified through the watermarking system.
More from Tech
Anthropic insists the process should not affect the quality of Claude’s responses. “To a reader, a watermarked response is indistinguishable from an unwatermarked one,” the company said.
The approach is also different from conventional AI-writing detectors, which typically analyse linguistic characteristics and look for patterns associated with machine-generated text. Anthropic said watermark detection instead checks for a specific signal embedded during generation.
That distinction matters because rewriting can affect whether the watermark survives. Anthropic expects minor changes to leave much of the signal intact, meaning light editing is unlikely to remove it completely. A substantial rewrite, however, can eliminate the watermark if essentially every word is replaced. The company argues that, in such a case, it becomes harder to describe the resulting text as the original AI-generated output.
The watermark’s presence will also depend on how Claude is used as an editor. If the model merely proofreads a long piece written by a person, there may be little newly generated material for the system to mark. Anthropic says the outcome will depend on the length of the material and the extent of Claude’s changes.
What happens to code and other files?
Code presents a different challenge. Unlike ordinary prose, software code often has to follow precise rules and produce a specific result. That leaves Claude with fewer opportunities to make interchangeable word choices while generating it.
As a result, Anthropic expects code to contain a much weaker watermark. There can still be opportunities to embed one where developers have discretion over wording, such as comments, but the company says the effect on the functional code itself should be negligible.
Anthropic is also using a separate mechanism for certain files rather than relying solely on text watermarking. It says digitally signed provenance information can help indicate whether a file has passed through Claude and whether its metadata has subsequently been changed.
The changes apply to Claude models launched from August 2, 2026, and are being rolled out across Anthropic’s products and services. The company says the measures form part of its commitments under Article 50(2) of the EU AI Act’s Code of Practice on Transparency of AI-Generated Content.
Anthropic also expects the change to extend beyond Claude. According to the company, other major AI model developers that have signed the same code of practice are expected to introduce their own watermarking systems.
The announcement has nevertheless prompted debate among Claude users, with some questioning what the technology means for privacy, authorship and the use of AI-assisted writing. Anthropic’s explanation makes clear that the watermark is not intended to prevent people from editing AI-generated text; rather, it is designed to provide a way of establishing whether generated material carries the model’s embedded signal.