Claude's replies and images now carry a mark you can't see, and EU regulators are the reason it exists.
According to a report, Anthropic has begun embedding invisible watermarks into everything Claude generates, covering both text responses and images. The trigger is the EU's AI Act, which pushes AI providers toward machine-detectable markers on synthetic content. Anthropic isn't limiting the change to European accounts. The watermarking applies to Claude outputs everywhere, regardless of where a user is logged in from. Anthropic has not detailed the watermarking method itself, so what's known so far comes from the report of the rollout rather than a technical write-up from the company.
This puts Anthropic alongside Google's SynthID and other provenance efforts, as major labs quietly build detection hooks into their default output rather than offering them as opt-in tools. It also shows how EU rules are shaping products worldwide. Anthropic could have geofenced the watermark to EU-only accounts and didn't, which says something about how it expects regulation to spread. For anyone relying on Claude text in contexts where origin matters, like classrooms, newsrooms, or court filings, that changes the calculus slightly.
Invisible watermarks have a mixed track record. Researchers have repeatedly shown they can be scrubbed with basic editing, so treat this as a compliance box being checked rather than a reliable way to catch AI-written text or images.