Anthropic Details Invisible Watermarking for Claude Text to Comply with EU Rules

Anthropic has shared technical details about how it will add invisible watermarks to text generated by its Claude AI assistant. The move is part of the company's compliance with the European Union's AI Act, which requires providers serving the EU market to mark AI-generated content so that it can be identified. Anthropic says the watermark will be rolled out globally at launch because it does not yet have a durable way to limit the feature by region. Future Claude models will generate watermarked text, while models released before August 2, 2026, are covered by a transition period and will gain watermarking over the coming months.

The watermarking approach is built on Google DeepMind's SynthID-Text method and works during text generation rather than after. AI language models produce text by repeatedly choosing the next word or word piece, called a token, from many possible options. Anthropic's watermark does not add hidden characters or alter a finished response. Instead, when Claude has several reasonable choices for the next token, it uses a secret key and some of the previous words as part of the randomness that decides which token to pick. To a reader, each individual choice looks normal, but across a long enough passage these choices leave a statistical pattern. A detector that has the key can check whether the sequence of words is consistent with what Claude would produce using that key and estimate the probability that Claude was involved in writing the text. Anthropic says internal testing found no impact on creativity, readability, or content quality, and that watermarking requires no extra tokens and has a negligible effect on generation speed.

There are important exceptions. For factual statements where only one answer is correct, the watermark does not interfere with the choice. The same applies to code, where swapping one term for another could break the output. Anthropic gives the example that after writing "2 + 2 =" there is a very clear best next token, and the watermark nudge is not applied. Code generally has less watermarking because exactness is often required, though parts of code with arbitrary choices, such as comments, can still be watermarked. This matches the SynthID-Text research, which notes that watermark detection works better when there is more entropy, meaning many equally good next-word options, and worse when the model's output is highly predictable. A translation produced by Claude carries a watermark because Claude chooses every word in the translated output. On the other hand, light proofreading of human-written text may leave too little Claude-generated material for reliable detection, since the watermark only applies to words Claude actually chooses.

Anthropic also plans to offer a watermark detection API, which is an application programming interface that lets other software check text. The API will estimate the likelihood that Claude was involved in writing a piece of text, but Anthropic stresses that this is not the same as proving who wrote it. The watermark cannot identify whether the text was written by another AI model, because other providers may use different methods and keys. A watermark can only indicate that Claude was likely involved at some point, and cannot distinguish "Claude wrote this" from "Claude heavily edited this." Light editing probably will not remove the watermark completely, but a full rewrite where every word is replaced will. Detection also becomes less reliable with short samples, as there are fewer word choices for the detector to analyze. For generated PNG, JPG, and SVG image files, Anthropic is taking a different approach: Claude will attach cryptographically signed C2PA provenance metadata indicating the file was created or processed with Claude, rather than modifying the image itself with an embedded watermark.

For website owners and businesses, invisible watermarking changes how AI-assisted content can be verified. As detection tools roll out, content teams may be able to confirm whether a text was likely generated or heavily edited by a particular AI model, which could matter for trust, compliance, and content integrity. AEU Hosting, which provides managed WordPress hosting secured end to end, can help website owners keep their publishing environment safe while they evaluate provenance tools and integrate new verification workflows.