A compliance-driven change for Claude

Anthropic has explained how it plans to add invisible watermarks to text generated by Claude. The company says Claude’s text marking system will use “a version of the SynthID-Text approach,” an open-source watermarking technology developed by Google DeepMind.

The move is tied to the European Union’s AI Act. According to Anthropic, the law’s transparency requirements call for synthetic or AI-manipulated audio, images, video, and text to carry machine-readable marks so that such content can be detected. In parallel, Anthropic says Claude-processed images will support C2PA, a standard used to attach verifiable provenance information to digital media.

How the watermark works

The watermark is not a visible label, nor does it insert obvious tags into Claude’s answers. Instead, it uses the way large language models choose words.

Anthropic gives a simple example: after the phrase “The weather today was cold and…,” the next word is very unlikely to be “sugary,” but words such as “overcast” or “grey” may both be plausible. In many cases, either choice would preserve the meaning of the sentence. Normally, a model may use a random number generator to decide among such low-stakes alternatives.

With watermarking, that randomness is changed. The system uses a key and a few preceding words to guide which acceptable word should be chosen. Across a longer passage, these small choices create a statistical pattern. Readers should not notice the pattern, but someone with the right key can detect it.

In plain terms, SynthID-Text-style watermarking creates a hidden statistical signature in generated writing by shaping probability-based word choices.

What Anthropic says about impact

Anthropic says the feature will not make Claude more expensive for users and will not have “any practical impact” on the quality or content of Claude’s outputs. That point matters because text watermarking can raise concerns that generated responses might become less natural or less useful.

The known facts from Anthropic’s explanation are limited but clear:

• Claude text will use a version of the SynthID-Text approach.
• Claude-processed images will support C2PA.
• Google Gemini has supported SynthID-Text since 2024.
• OpenAI has not detailed a ChatGPT text watermarking plan in its AI Act compliance roadmap, though it will also be subject to the law’s requirements.

Industry direction

Claude is not the only AI product affected by the EU’s transparency rules. The broader direction is that major AI systems will need to make synthetic content easier to identify, even when the signal is invisible to end users.

The important open questions are operational rather than conceptual: who gets access to detection keys, how detection will work across languages and edited text, and how reliable the signal will be in real-world sharing. Anthropic has not provided those details in the material available.

The broader trend is clear: AI content provenance is becoming part of the default infrastructure for generative AI. Model quality and price will still matter, but compliance, traceability, and machine-readable disclosure are now becoming core product requirements.