Claude's 'AI-written' label vanishes once you change a setting

A routine setting

Take a hypothetical case. A startup sells shopkeepers an assistant, under its own brand, that drafts replies to customer reviews. It runs on Claude Sonnet 4.6. The developer lowered the temperature, a common tweak, so the replies stay consistent from one day to the next. Nothing in the code mentions Europe.

According to an email Anthropic has just sent to its API customers, that one setting is enough to strip these texts of the invisible mark the EU AI Act calls for. The trade site The Decoder, which saw the email, quotes the instruction: anyone counting on the watermark for their own compliance should leave those settings at their defaults.

It's a short sentence. It says a lot about how the European obligation gets split up.

What Anthropic announced

According to The Decoder, Anthropic is extending its text watermark to four older models: Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. Newer models already carry it.

On Opus 4.6 and Sonnet 4.6, the two models at issue, the mark would only be applied if three parameters, temperature, top_p and top_k, stay at their defaults. Changing a single one is enough to skip it, the outlet sums up.

That condition shows up nowhere else. Anthropic's help page on watermarking doesn't mention it, nor does the blog post explaining the watermark, nor does the API documentation. As of November 1, the help page doesn't even tick a text watermark for any of the four models yet: the announcement came before the page caught up.

For the principle, we covered it when it was announced in August: a statistical signature woven into word choice, invisible, applied at the model level and worldwide. Customers can't turn it off, The Decoder notes. Except, it turns out, through the side door.

Three dials on chance

A model writes one word at a time. At each step it draws the next word from a pool of plausible candidates. Temperature sets how much chance goes into that draw; top_p and top_k set how big the pool is. Anthropic explains that its watermark changes only the source of that randomness: instead of an ordinary die, a secret key.

Why changing the dials cuts the mark on these two models, the company doesn't say. Its documentation does show that the question can only arise there. On Mythos 5, on Opus 4.7 and on every later model, a non-default value returns an error: nobody can touch the die anymore. In the rollout list, Opus 4.6 and Sonnet 4.6 are the only two where the dial still moves.

And it's on its way out. The new API version, dated 2026-10-21, removes all three parameters on every model. Anthropic even describes our startup's exact case: on Sonnet 4.6, a request that lowered the temperature for steadier output loses that effect, and nothing replaces it.

Customers who rely on it can stay on the old version. So the gap only lives in integrations left behind, which makes for a narrow window, but not an empty one.

Two floors of obligation

This is where the AI Act comes in. Its Article 50 requires providers of generative AI systems to mark their output in a machine-readable way.

But the European Commission is clear about who counts as a provider: whoever puts a system on the market under its own name, whether or not it is established in the Union. The startup selling a Claude-based assistant isn't just a customer, then. By that definition, it looks a lot like the provider of its own system.

Anthropic says as much in its own way. Its help page tells those who deploy Claude in their products to assess for themselves what Article 50 requires of them. And the instruction in the email talks about their compliance, not Anthropic's.

The Commission saw this split coming. Its FAQ mentions techniques a model provider can put in place to help the providers of downstream systems comply. For Anthropic's customers, the watermark plays that role: a service rendered to the floor below, which only works if the floor below doesn't unplug it.

So the building has two floors. The landlord wires the alarm into the walls, and the tenant finds out by mail that a switch in his kitchen turns it off.

The calendar

The Decoder writes that Anthropic cites the AI Act's December deadline to explain its timing. The Commission is the one that gives the date: Article 50 has applied since August 2, and systems placed on the market before that date have until December 2, 2026 to meet the marking obligation, and that obligation only. Our explainer on the August 2 deadline linked this grace period to the digital Omnibus, Regulation 2026/1744, published in the Official Journal on July 24.

A system launched after August 2 gets no grace period at all. For our startup, then, the date it put its assistant on sale matters as much as the model it uses. Fines can reach 15 million euros or 3% of worldwide annual turnover.

What we don't know

Several things remain open. First, the exact text of the email: we only know it through The Decoder. Then there's how a supervisory authority would read the case of an integrator who set a temperature without knowing what it switched off. None has ruled, and we won't settle here a legal question nobody has yet decided.

Checking for yourself isn't easy either. Anthropic's detection tool is in private preview. It's open to regulators, media and researchers, and, on request, to companies that need to verify the marking for their own compliance.

At OpenAI, the philosophy runs the other way. According to The Decoder, its textGrain watermark is on by default in ChatGPT and Codex for European users, but off on the API until the customer turns it on. Anthropic imposes the mark and lets a setting switch it off; OpenAI ships it switched off on the API. Either way, the switch ends up in the customer's hands.

We saw the same thing with the identifier Microsoft slips into Paint images: every provider invents its own way of marking. The regulation, for its part, says who answers.

An instruction at the bottom of the page

What stands out most is the format. A condition for complying with European law arrives in an announcement email, and nowhere in the documentation a developer reads while writing code.

For the team that set its temperature months ago, nothing has changed on screen. Its replies go out, steady, exactly as before. That's precisely what's missing: a sign saying that, on two models, the mark got left in the cloakroom.