Anthropic is finally explaining how it plans to watermark Claude’s output, and the details are more nuanced than the panic on Reddit and X suggests. The company published a technical explainer Friday confirming it will use Google DeepMind’s SynthID-Text approach, which embeds detectable patterns in word choices that are invisible to human readers but can be verified by anyone holding the right key. The trigger for all this? The EU AI Act’s Transparency Code, which requires AI developers to make generated content identifiable.
The company insists the watermark won’t degrade output quality. When Claude faces a low-stakes choice between two valid words, it can quietly encode a signal. But code gets special treatment. Anthropic admits there’s less room for watermarking in programming because the model needs to produce working code, not arbitrary word selections. The watermark will show up mainly in code comments where word choice is flexible. ‘It will have a negligible effect on the actual code produced,’ the company said.
What about someone who runs watermarked text through an editor or rewrites it entirely? Anthropic says light editing probably won’t strip the watermark completely, but a full rewrite where every word changes will. In that case, the company adds, it’s debatable whether the text can even still be called AI-generated. The detection API Anthropic plans to release is fundamentally different from services like Pangram that hunt for stylistic tells—phrases like ‘this isn’t X, it’s Y’—rather than checking for an actual embedded signal.
Anthropic also made clear Claude won’t be alone in carrying watermarks. Other major model developers have signed the same Code of Practice and will roll out their own implementations. That’s significant for anyone who assumed switching chatbots would sidestep the issue. The user backlash has been real—Business Insider reported dozens of X users claiming they’ve canceled subscriptions—but the technical reality is that most people won’t notice the watermark at all. The people who should care most are those trying to pass Claude output off as human, and that’s precisely the point.
