Anthropic embeds text watermarks in Claude outputs to comply with EU AI Act, raising questions for law firms about disclosure and document integrity

Anthropic has announced that its Claude large language model will embed invisible text watermarks in AI-generated outputs, a move driven by requirements under the EU AI Act that mandate identification of AI-generated content for all LLM providers operating in the EU. The watermarking mechanism works by exploiting low-stakes word choices during text generation: where multiple plausible words exist, Claude selects based on a secret key rather than a random number, leaving a pattern undetectable to a human reader but verifiable by anyone holding the key. Anthropic has confirmed it is developing a watermark detection API to allow verification of whether content was Claude-generated, with implementation details still being worked out. The disclosure that a watermark exists does not affect ownership or authorship of an output: Anthropic has stated explicitly that a watermark does not change a user's rights under its terms or indicate who is legally responsible for the content. For lawyers, the principal risk is not the watermark itself but the question of whether a court or client could detect that a document was AI-assisted and what disclosure obligations follow. Firms using open-weight models that have not yet incorporated watermarking systems face a compliance gap if those outputs are used in the EU. The system also does not address hallucinations: a watermark signals production method, not accuracy.