Claude Is Getting Invisible SynthID Watermarks to Comply With the EU AI Act

Anthropic will embed undetectable word-pattern watermarks in Claude's text — a version of Google DeepMind's SynthID-Text — to satisfy Europe's new AI transparency rules that took effect August 2.

Every response Claude writes will soon carry an invisible signature. Anthropic has announced that future Claude models will embed a text watermark — a subtle, machine-readable pattern woven into word choices — to comply with the European Union's AI Act, whose transparency rules for marking AI-generated content took effect on August 2.

The technique is "a version of the SynthID-Text approach," the open-source watermarking method developed by Google DeepMind. It works by nudging the model's low-stakes decisions: when Claude finishes the sentence "The weather today was cold and…," either "overcast" or "grey" would do, and normally the choice falls to a random number. Watermarking swaps in a keyed source of randomness instead, so the sequence of words forms a pattern a reader cannot see but a detector holding the key can flag.

Anthropic stresses the marks are undetectable to humans, add no hidden characters and no extra cost, and carry no identifying information — they cannot be traced to a specific person, organization, or chat. The company is also adding C2PA support for Claude-processed images. The same EU rules mean Claude won't be the only model bearing watermarks: Google's Gemini has supported SynthID Text since 2024, while OpenAI has yet to detail text-watermarking plans for ChatGPT but is subject to the same requirements.

Sources: Anthropic, The Verge, The Hindu