Anthropic details Claude's SynthID-Text watermarks for EU AI Act compliance
Anthropic has clarified how it plans to apply invisible watermarks to Claude generated text in order to comply with Europe's AI transparency rules. On Friday, Anthropic announced that Claude's text marking system is a version of the SynthID-Text approach, an open-source watermarking technology developed by Google DeepMind that creates detectable patterns using wording probabilities. The watermarking feature, alongside C2PA support for Claude processed images, is being introduced to meet Anthropic's obligations under the European Union's AI Act, which requires synthetic audio, image, video, and text to include machine-readable marks so the content can be detected as artificially generated or manipulated. Anthropic says the text watermarks will not make Claude more expensive for users, or have any practical impact on the quality or content of Claude's outputs.
Anthropic explained the mechanism with an example: in a sentence like "The weather today was cold and...", the next word is very unlikely to be "sugary" but is quite likely to be "overcast" or "grey". Under most circumstances it does not matter much to the reader which of those two similar words the model picks, so ordinarily the choice is settled by a random number. Watermarking uses these low stakes word choices, which occur many times over a generated piece of text, to leave a pattern in Claude's responses. The pattern is undetectable to a reader but detectable to anyone holding a key that encodes it. With watermarking on, the choice is still effectively random, but the source of that randomness changes: instead of an arbitrary random number generator picking the next word, the system uses the key together with the preceding few words to settle which word the model picks.
Anthropic noted that the EU's AI transparency requirements also apply to other major AI developers, so Claude will not be the only model carrying text watermarks. Google's Gemini chatbot has supported the SynthID-Text solution since 2024. OpenAI has not detailed any text watermarking plans for ChatGPT in its AI Act compliance roadmap, though it will also be subject to the law's requirements.
Key facts
- Anthropic's new Claude text watermarking system is described as a version of Google DeepMind's open-source SynthID-Text approach.
- The feature, along with C2PA support for Claude-processed images, is meant to satisfy the EU AI Act's requirement that synthetic audio, image, video and text carry machine-readable marks identifying them as AI generated.
- Anthropic says the watermark will not raise Claude's cost for users or affect the quality or content of its outputs.
- The watermark works by steering low-stakes word choices (where multiple words fit equally well) using a key and preceding words instead of pure randomness, leaving a pattern invisible to readers but detectable with the key.
- Google's Gemini has used SynthID-Text since 2024, while OpenAI has not detailed text watermarking plans for ChatGPT despite also being subject to the EU AI Act.
Why it matters
This is Anthropic's first detailed explanation of how it will meet a concrete, binding requirement of the EU AI Act: that synthetic text, alongside audio, image and video, carry a machine-readable mark showing it was AI generated. Rather than building an entirely new system, Anthropic is adopting a version of Google DeepMind's existing open-source SynthID-Text approach, which signals a degree of convergence among major labs on how to solve text watermarking rather than each building a proprietary, incompatible scheme.
Who it affects
It affects anyone using Claude to generate text once the watermarking is in place, since the marks are added to Claude's outputs. It also affects other large AI developers operating in the EU: Google's Gemini already uses SynthID-Text, and OpenAI's ChatGPT will be subject to the same AI Act requirement even though the article notes OpenAI has not detailed its own text watermarking plans.
How to use it
There is nothing for a user to configure: the watermark is applied automatically to Claude's text output as part of Anthropic's AI Act compliance work, alongside C2PA support for Claude-processed images. Anthropic states the feature carries no extra cost and no practical effect on output quality or content, so usage of Claude itself does not change.
How solid is it
The explanation comes directly from Anthropic and describes a known, previously published technique: SynthID-Text is Google DeepMind's open-source watermarking method, already deployed in Gemini since 2024, which lends the approach some track record. That said, the source gives no accuracy, false-positive or detection-rate figures for Claude's implementation, no detail on how the watermarking key is generated, stored or distributed to detectors, and no confirmation of whether the feature covers all Claude models and tiers or only some.
Risks and caveats
Key technical and rollout details are unstated: there is no specific launch date for when the watermarking goes live beyond the Friday announcement, no stated deadline for the EU AI Act's transparency requirements, and no word on which Claude models or subscription tiers will carry the mark. The source also does not name which other competitors besides Google and OpenAI have or have not adopted SynthID-Text, so the broader competitive picture is incomplete.
“have any practical impact on the quality or content of Claude's outputs”
— Anthropic