Europe Now Requires AI Watermarks — But Nobody Can Verify Them
Invisible tags now mark AI-written text — and no outsider can check if they work.
On August 2, 2026, a European rule took effect: any company selling AI that writes text must secretly tag what it produces, so AI writing can be told apart from writing by a person. Days later, Anthropic revealed that every Claude model released after that date adds an invisible watermark to everything it writes — turned on by default, with no way for users to switch it off. Google's Gemini has done the same since 2024.
Users pushed back hard. Some said the tags make answers worse, especially computer code. Others worried the watermark quietly hides information about who typed the prompt. Oddly, some complained the tag is easy to strip out, while others said it can never be escaped. Anthropic answered that quality is unchanged, no identifying information is hidden, and light editing won't remove the tag. The researchers' key point: right now, none of these claims — on either side — can actually be checked by an outsider.
Because no public tool can test the real, deployed systems, the team ran the open-source version of the watermark on two freely available AI models instead. In ordinary prose, the watermark changed the writing no more than simply re-rolling the dice on a random setting. In code, it cost roughly three points of correctness on one model and was too small to measure on the other. Most striking of all: the tag was barely detectable — close to a coin flip.
The authors argue the real failure isn't watermarking itself, it's that nobody can verify it. They call for matched sample outputs, published settings, independent audits, a shared testing standard, and detection tools that work across different companies. For you, this matters because the rules now promise you can spot AI text, yet the evidence stays locked inside the companies that wrote it. Treat "this is AI-generated" claims — and "this is human" ones — as unproven for now.
- Since August 2026, European law requires AI companies to invisibly tag AI-written text — Claude and Gemini both do it automatically, with no off switch.
- Testing the open-source version, researchers found the tag barely changed normal writing, cost about 3 points of accuracy on code in one model, and was almost impossible to detect.
- The study's real complaint: there is no independent way to check whether watermarks work or hurt quality, so both users' fears and companies' promises stay unproven.
Why It Matters
Rules now promise you can tell AI text from human text, but the proof sits behind company doors.