Lun. Nov 16th, 2026
A magnifying glass reveals a circular watermark on a blank document as its edge dissolves into particles, illustrating the limits of detecting AI-generated text.

Anthropic has announced an invisible watermark for the text Claude generates, starting with its upcoming models; the ones already in service will follow over the coming months. The reason is not a sudden passion for transparency: it is the EU AI Act, plus the Code of Practice on AI-generated content that around 190 organizations signed in July 2026. The watermark applies globally from the start, because carving out Europe was apparently harder than watermarking the whole planet. Nothing says “proportionate regulation” like a European rule that instantly governs a user in Ohio.

The mechanism itself is genuinely clever. It is SynthID-Text, the method Google DeepMind published in Nature in 2024, building on an idea Scott Aaronson floated in 2022. When the model must choose between equally good words, the choice is still random; what changes is where the randomness comes from. Instead of pure chance, it follows a pattern derived from a secret key. The text looks just as random to any reader; whoever holds the key can check whether the pattern is there. Anthropic’s own analogy is a Monopoly game where the dice are replaced by the digits of pi. The company underlines three guarantees: quality is unchanged in internal testing, in creativity and readability alike; there are no hidden characters, no extra tokens, and no penalty on speed or price; and nothing in the watermark traces back to a user or an organization.

Now read the fine print, which to Anthropic’s credit is printed in their own announcement. The watermark answers exactly one question: was Claude probably involved in producing this text? It does not prove a text is human. It does not recognize other models. It works poorly on short texts, is sparse in factual passages full of names and dates, and is thinner still in code, where arbitrary wording choices, comments included, are scarcer. And editing matters: a light touch-up probably will not strip the mark entirely, but rewriting every word will.

Put those limits together and you get the real picture. Anyone who wants to hide has an exit printed in the manual: rewrite the thing completely and the stamp is gone. That takes effort, but it is exactly the effort a determined cheat or a propaganda operation can budget for as a production step. Meanwhile the honest professional who openly uses Claude to draft reports is the one most likely to be wearing the stamp. We have built a metal detector that beeps at keychains and leaves a marked service door around the back. And be clear about what the stamp even means: AI-involved, not false. A watermarked text can be perfectly accurate; a clean one can be a lie written entirely by hand.

And who gets to check? Not you. Detection runs through an API operated by Anthropic, initially in private preview for eligible parties: regulators, law enforcement, media, fact checkers, researchers, educational institutions, European civil society, and companies that are themselves under compliance obligations. Transparency about machine-generated content is now an invitation-only privilege, complete with a registration form. The ordinary reader, the person this whole apparatus supposedly protects, still has to guess.

For compliance functions the implication is concrete and slightly uncomfortable. Public documents and communications your organization produces with Claude may be flagged, probabilistically, as AI-involved by whoever holds the key, and longer texts more reliably than short ones. The reverse is not true: a clean result proves nothing about human origin. As evidence, the watermark is a one-way street with a hole in the middle.

The mathematics is elegant. The engineering is sound. But as a compliance instrument this is a probabilistic stamp that certifies little, fades with handling, and can only be read with glasses Anthropic lends to those it deems eligible. Brussels will say the law is applied, Anthropic will say it is compliant, and anyone actually trying to tell human from generated will be standing almost exactly where they stood before. Regulation achieved; problem intact.

Raffaele Di Marzio

All my “insane” books on cybersecurity and governance are here 👇 https://cyberium.limited/bookshelf.html