Anthropic explains how Claude’s invisible text watermarks will work

Anthropic announced on Friday that it will watermark text generated by its Claude AI model using a version of Google DeepMind's open-source SynthID-Text technology, to comply with the European Union's AI Act transparency requirements. The watermarking will not increase costs or affect output quality, and it will be detectable only with a key. Google's Gemini has supported SynthID Text since 2024, while OpenAI has not detailed text watermarking plans for ChatGPT.

Anthropic has clarified how it’s planning to apply invisible watermarks to Claude-generated text in order to comply with Europe’s AI transparency rules. On Friday, Anthropic announced https://www.anthropic.com/news/claude-text-watermark that Claude’s text marking system is “a version of the SynthID-Text approach” — an open-source watermarking technology developed by Google DeepMind /2024/10/23/24277873/google-artificial-intelligence-synthid-watermarking-open-source that creates detectable patterns using wording probabilities. Anthropic explains how Claude’s invisible text watermarks will work It’s using ‘a version’ of the open-source SynthID-Text system Google developed. It’s using ‘a version’ of the open-source SynthID-Text system Google developed. This watermarking feature, alongside C2PA support /ai-artificial-intelligence/977823/anthropic-claude-ai-watermarks-c2pa-text-images for Claude-processed images, is being introduced to meet Anthropic’s obligations under the European Union’s AI Act /ai-artificial-intelligence/974571/eu-ai-act-transparency-labels-rules-deepfakes , which requires synthetic audio, image, video, and text to include machine-readable marks that enable the content to be detected as artificially generated or manipulated. Anthropic says the text watermarks won’t make Claude more expensive for users, or “have any practical impact on the quality or content of Claude’s outputs.” Here’s Anthropic’s explanation for how it works: Take the sentence “The weather today was cold and…”. The next word is very unlikely to be “sugary.” But it is quite likely to be “overcast” or “grey.” Under most circumstances, it doesn’t matter much to the reader which of these latter two words the model ultimately chooses—the meaning of the sentence is largely the same either way. In cases like this, the choice is settled by a random number. Watermarking uses low-stakes choices like these—which occur many times over a piece of generated text—to leave a pattern in Claude’s responses. That pattern is undetectable to the reader, but is detectable to anyone who has a key that encodes it. When watermarking is used, choices are still made at random, but the source of the randomness is different. Instead of using an arbitrary random number generator to pick the next word, watermarking uses the key and a few words that come before to settle what word the model should pick. As Anthropic notes, the EU’s AI transparency requirements also impact other major AI developers, so Claude won’t be the only model introducing text watermarks. Google’s Gemini chatbot has supported the SynthID Text solution since 2024 https://deepmind.google/blog/watermarking-ai-generated-text-and-video-with-synthid/ , and while OpenAI hasn’t detailed any text watermarking plans https://help.openai.com/en/articles/8912793-provenance-signals-content-credentials-synthid-in-openai-generated-content for ChatGPT in its AI Act compliance roadmap https://help.openai.com/en/articles/12141645-eu-ai-act-openai-resources-and-customer-guidance , it will also be subject to the law’s requirements. Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.