A study says Claude's watermark barely marks anything

Anthropic started embedding a SynthID-Text watermark in every Claude model released after August 2, 2026, when EU AI Act Article 50 took effect. There's no opt-out. arXiv 2609.09604 tries to check whether the watermark actually works, and can't test the real thing at all.

No public tool exists to probe Anthropic's deployed watermark. So the authors ran the open-source SynthID-Text implementation on two open-weight models instead, since that's the closest stand-in anyone outside Anthropic can build. On prose generation, the watermark's effect on quality didn't exceed what you'd get from just changing the sampling seed. On code, it cost three points of correctness on one model and fell below measurement on the other. Detection accuracy, the entire point of a watermark, stayed near chance in their tests.

That's a rough result for a mechanism regulators are treating as a working disclosure tool. If detection can't reliably tell watermarked text from unwatermarked text, the watermark isn't doing the job the mandate assumes it's doing.

The authors' argument isn't that the watermark is badly built. It's that nobody outside Anthropic can check, and that gap is the actual governance failure. A regulation that requires watermarking but leaves verification to the same company doing the watermarking doesn't give outsiders anything to audit.

What the paper doesn't have: results from Anthropic's actual production models, since those aren't testable by outsiders. Everything here is inference from an open-source stand-in on different models. That's a real limitation, and it's also the whole point, since a compliance regime built around something researchers can't independently test is a regime built on trust rather than proof.