Anthropic releases invisible text watermarks for Claude AI in EU compliance move

Anthropic has unveiled details about how it will embed hidden watermarks into text generated by Claude, addressing Europe's strict requirements for tracking artificially created content. The system relies on a variant of SynthID-Text, an open-source watermarking approach created by Google DeepMind.

How the watermarking works#

The approach exploits the freedom that language models have when generating text. When Claude writes a sentence like "The weather today was cold and…", the next word could reasonably be "overcast" or "grey"—both make sense and convey similar meaning to readers. Rather than letting pure chance determine which word appears, the watermarking system uses a cryptographic key to steer that choice. This pattern repeats throughout the text, creating an invisible fingerprint that machines can detect but humans won't notice.

According to Anthropic, "the choice is settled by a random number" in normal operation, but with watermarking enabled, that randomness comes from a seeded key instead, making the output pattern verifiable.

EU regulatory backdrop#

The watermarking requirement stems from the European Union's AI Act, which mandates that synthetic audio, images, video, and text all include machine-readable markers identifying them as artificially generated or modified content. Anthropic says the feature won't add cost for users or degrade Claude's performance or output quality.

Industry-wide adoption#

Claude isn't alone in this shift. Google's Gemini chatbot has supported SynthID-Text since 2024, and OpenAI's ChatGPT will eventually face the same legal obligations, though the company has yet to announce specific text watermarking plans. Anthropic is also rolling out C2PA support for Claude-processed images as part of the same compliance effort.