Anthropic's new watermark is turned on for every Claude reply, and a fresh study says no outsider can verify if it's doing what the company claims.
The EU AI Act's Article 50 rules kicked in on August 2, 2026, requiring generative AI providers to mark their output so it can be detected as machine-made. Days later, Anthropic said every Claude model released since then embeds a SynthID-Text based watermark in all generated text, on by default with no opt-out. Google has run the same SynthID-Text system in Gemini since 2024. Users pushed back, arguing the watermark hurts output quality, especially for code, secretly encodes identifying details, and is somehow both trivial to strip and impossible to escape; Anthropic denied all three.
The new paper's point isn't that watermarking is bad. It's that neither side's claims can currently be checked, because no public tool tests the actual deployed Claude or Gemini systems. Running the open-source SynthID-Text code on open-weight models instead, the researchers found the prose impact was no bigger than switching the random seed, the code-quality cost topped out at three points of correctness, and detection was barely better than a coin flip.
That's the real story here: a watermarking law arrived years before any shared, auditable way to check whether a watermark works, so the compliance box got ticked well ahead of anyone's ability to grade the homework.