Anthropic adds text watermarking to Claude for EU compliance

Anthropic announced it will implement invisible watermarking in its Claude chatbot to comply with the European Union AI Act. The company stated the technique modifies token selection probabilities during generation to embed detectable signals without altering visible content. Critics warn that constraining model sampling could degrade output quality and exacerbate existing stylistic quirks associated with large language models. Anthropic maintains the change is necessary to satisfy upcoming EU transparency requirements for synthetic content detection. This implementation represents one of the first major technical adaptations by a frontier lab specifically targeting EU regulatory mandates. Industry observers are monitoring whether watermarking impacts benchmark performance or user satisfaction metrics. The move signals growing operational friction between open-ended generative capabilities and emerging government oversight frameworks. Compliance costs and technical trade-offs may influence how other providers approach similar international regulations.