Anthropic's invisible Claude watermark leaves detection details unanswered

Anthropic has rolled out a watermarking system for content generated by Claude, its AI assistant, to comply with EU transparency rules. However, the company has yet to release technical details about how the watermarks can be detected or share testing results on false positive and negative rates—raising questions about whether the approach will actually meet regulatory expectations.

EU pushes for AI transparency#

The European Union's AI Act mandates that AI systems disclose their involvement in generating content to preserve "the integrity and trust in the information ecosystem." The requirement aims to combat misinformation, fraud, and impersonation at scale by ensuring people "know when they are interacting with AI or exposed to AI-generated content," according to the European Commission's guidance.

Yet there's an inherent tension in Anthropic's approach: an article that is entirely AI-generated and never reviewed by an editor receives a watermark, but may not get a visible, reader-facing label if a human editor properly reviews it before publication.

The detection problem#

Anthropic has not specified a timeline for releasing detection methods or provided evidence that the watermarking system will reliably identify AI-generated material without triggering false alarms. The EU expects AI firms themselves to develop solutions, given how rapidly the technology evolves and the ongoing need to "trace the origin of information."

The regulatory framework, however, leaves implementation details largely to companies. The EU only stipulates that "techniques and methods should be sufficiently reliable, interoperable, effective and robust as far as this is technically feasible"—a standard that remains open to interpretation.

High stakes enforcement#

Failure to comply carries serious consequences. Under the AI Act, violations can result in fines up to 15 million euros or 3 percent of a company's global annual revenue. Anthropic has committed to improving both its watermarks and detection methods to meet EU requirements, but the gap between announcement and demonstrated capability remains wide.