They signed the words

Nearly every big AI lab now hides an invisible mark in the text it writes, but the mark can say much less than people think.

On Friday evening a marketing executive called Pete Sena posted a screenshot of an email from Anthropic with the line "Claude gonna start snitching on Tuesday." He explained that from 3 November, text from Claude carries an invisible watermark. You can't see it, it survives copy and paste, and it can survive some editing. It doesn't say who you are, he wrote, only that Claude touched the words, and even a quick proofread counts. "If your brand voice runs on Claude, write like someone might check. Because now someone can." By this morning the post had been seen a little over four million times and bookmarked nearly three thousand times, which tells you that a lot of people read it and went quietly to look at how much of their own week Claude had written.

He is mostly right, and the parts where he's slightly wrong are the interesting bit. So this is a piece about watermarks, and about the odd week in which nearly every big lab ended up marking its text, and the reasons were all the same, and the caveats were nearly all the same, and the one company that held out is the one you'd least expect.

What actually changes tomorrow

Anthropic's email went to API customers last week. It said that Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6 would start carrying the same watermark that its newer models, Opus 5.5 among them, have carried since launch. Bedrock may take a few extra days. The company calls it an "imperceptible pattern in word choice" that doesn't change meaning, quality or readability, adds no characters or tokens, doesn't change the price, and "contains no information about the user, their organization, or their conversations with Claude."

That last sentence is the one Pete Sena's post gets right and a lot of the replies underneath it get wrong. There's no serial number in the text and nothing that points back to an account. What the mark says, if a detector finds it, is that a Claude model probably chose these words. It doesn't say which customer asked, and it doesn't say how much a person changed afterwards.

The mark is applied at the model level, so it turns up wherever the model runs. Anthropic says that includes AWS, Google Cloud and Microsoft Foundry, as well as the Claude apps, Claude Code and the API. Customers can't turn it off. And Anthropic is applying it worldwide rather than only in Europe, for a reason it was refreshingly honest about in its own explainer: "we don't yet have a durable way to scope it by region."

Then there is the loophole. On Opus 4.6 and Sonnet 4.6, the two oldest models on the list, Anthropic says the watermark only applies when developers leave the three sampling settings, temperature, top_p and top_k, at their defaults. Those are the dials that control how much randomness the model uses when it picks the next word. Change any one of them and the mark doesn't go on. The email doesn't hide this. It says, plainly, "If you rely on the watermark for your own compliance, keep these settings at their defaults." The Decoder, which broke the detail on Saturday, pointed out the awkward part, which is that a lot of people still think Opus 4.6 is Anthropic's best writing model, better than Opus 5.5. So the one model a writer might most want to use without a mark is also the one where a single setting removes it.

I don't think that's a scandal. It looks like an engineering limit on older weights rather than a wink, and Anthropic flagged it itself. But it's worth holding on to, because it tells you exactly what a watermark is, which is a property of how text was sampled, not a property of the text.

Next Saturday, Grok

On Saturday xAI, which its own developer docs now call SpaceXSI, published a page called Output Watermarking. From 7 November, text from seven Grok models on the API carries an invisible watermark by default: grok-4.7, 4.6, 4.5 and 4.3, and the three grok-4.20 variants, the reasoning, non-reasoning and multi-agent ones. It covers ordinary replies, structured outputs and tool-call arguments, on the Responses API and the Chat Completions API, on the global endpoint and the regional ones. There's a new response header, `x-watermark`, which comes back true when an output was marked. The reason given is one sentence long: "SpaceXSI watermarks text model output to comply with the EU AI Act."

The details are a little different from Anthropic's. Team admins can switch the mark off per model in the console, but only after confirming that "an exception under applicable law applies to your use, and that you are solely responsible for any marking, disclosure, or other legal requirement." Every opt-out lands in the team's audit log. Some models may be "locked on" and can't be turned off at all. Detection is statistical, it comes back as watermarked, not watermarked or uncertain, and access to the detector is by application only. The page says quality and speed are unchanged. It gives no accuracy figures and no false-positive rate.

The consumer Grok apps aren't part of it. Grok's own account spent much of the weekend answering that question under other people's posts, saying that Grok on X and in the apps has no text watermark planned, and that the API change is only for the listed models.

What makes this the strange part of the story is that xAI was the holdout. Andrew Curran put it neatly on Saturday night: "xAI was the last holdout, so the only model left that (as far as I know) does not watermark text is ChatGPT, and only ChatGPT in North America." Anthropic, OpenAI, Google, Meta, Microsoft and Mistral all signed the EU's Code of Practice on Transparency of AI-Generated Content over the summer. Anthropic says around 190 organisations signed it in total. xAI didn't. In August it stayed out of the voluntary code altogether. And now, three months later, it's watermarking by default anyway, with an opt-out that asks you to put your legal exposure in writing.

That's the clearest sign I've seen that the code wasn't really the thing doing the work. The law was.

The law and its December date

Article 50 of the EU AI Act is the transparency article. Since 2 August it has required the providers of generative systems to make sure their outputs, text included, are "marked in a machine-readable format and detectable as artificially generated or manipulated." The code of practice is voluntary. The article isn't. Signing the code is just the easiest way to show you're complying with it.

There's one grace period, and it's the reason this week is busy. In July the EU's Digital Omnibus gave generative systems that were already on the market before 2 August an extra four months to meet the marking duty, until 2 December 2026. That's the deadline Anthropic named in its email, and it's why the models being retrofitted now are the older ones. Anything launched after August had to ship marked. Anything launched before has a month left.

The Commission's guidance also explains the shape of the exceptions that keep turning up in the company docs. Source code is out of scope. So are short sequences of letters or numbers, and outputs that go from machine to machine without a person ever seeing them. The code of practice doesn't require marks on text shorter than 200 tokens, about 150 words of English, which is roughly the point below which none of these methods work reliably anyway. And the guidance lets a company building on top of a model lean on the model provider's marking, so long as both marking and detection are actually available. That last clause matters more than it sounds, because it's why the big labs had to do this at the model level. If they hadn't, every startup reselling Claude or Grok in Europe would have had to build its own.

How the mark works

All of these methods, Google's SynthID, OpenAI's textGrain, Anthropic's (which The Decoder reports is a variant of SynthID) and now xAI's, work in roughly the same way. When a language model writes, it doesn't pick the single most likely next word. It looks at a spread of candidates, each with a probability, and samples one. A watermark tilts that sampling very slightly, using a secret key, so that over hundreds of words the choices fall into a pattern that ordinary writing wouldn't produce by chance. A detector that knows the key can score a passage and say how unlikely its word choices would be without the tilt. Nothing is added to the text. The pattern is the choices themselves.

Once you see it that way, most of the weekend's questions answer themselves.

It's why the mark survives copy and paste. There's no hidden character to strip, because the words are the mark.

It's why it fades with heavy rewriting or translation. Google DeepMind has said this about SynthID since 2024: confidence "can be greatly reduced when an AI-generated text is thoroughly rewritten or translated to another language." Swap enough words and the pattern drowns.

It's why it's weak on factual answers. If the question is the capital of France, there's only one right word and no room to tilt the choice without getting it wrong. Watermarks live in the slack, so they're strongest in essays, emails and marketing copy, which happens to be exactly what Pete Sena's audience was worried about.

It's why Anthropic's loophole exists. If the mark is applied by shaping the sampling, and the older models only had it built for the default sampling, then changing the sampling settings goes around it.

And it's why the spelling question that ran through the replies under Grok's posts has a slightly unsatisfying answer. Someone asked whether a document a person wrote entirely, where the AI only fixed spellings, could carry a mark. Grok's account answered, fairly, that a dictionary spell-checker has no sampling step and so leaves no mark, while a language model correcting text still generates it a token at a time and can leave a pattern. Whether a few corrections are enough for a detector to say anything is another matter. A light proofread probably leaves too few choices to score. A model that "tidies up" a paragraph by rewriting half of it leaves plenty.

What the mark can't tell you

OpenAI, which has spent longer than anyone being cautious in public about this, wrote the clearest list of what a text watermark doesn't do when it launched textGrain in October. It doesn't measure human contribution. It doesn't establish ownership or responsibility. It doesn't identify the user. It doesn't verify accuracy. And "the absence of a detected watermark does not prove human authorship," because text can be too short, edited, translated, made by an older model, or made by another company's tools.

That last point is the one I'd underline for anyone who teaches, edits or hires. The weekend's version of the story, in a lot of posts, was that there's now a way to tell whether something was written by AI. There isn't. There's a way for a small number of approved organisations to tell whether something was probably written by one particular company's models, as long as it's long enough and hasn't been reworked much. Every lab is a separate detector with a separate key. Anthropic's detection API is in private preview, for regulators, law enforcement, media, fact-checkers, researchers, educational organisations, EU civil society groups and companies with their own compliance duties. OpenAI is handing out access case by case. xAI wants an application form. Only OpenAI has published test numbers, aiming to wrongly flag about one unmarked text in a hundred, and even so it said outright that it isn't making its detector public because of "the risk of missed watermarks and false positives."

So the honest picture is a dozen locks with a dozen keys, mostly held by the companies themselves, and a few open-weight models outside the system entirely. A watermark is good at answering one narrow question for an investigator: did this batch of spam, this flood of fake reviews, this run of scam emails come off our servers? That's a real use, and at scale it may be the main one. It's much worse at the question a worried teacher is actually asking.

There are costs on the other side, too. John Gruber has argued that because no two synonyms mean quite the same thing, choosing words partly by a key rather than purely by fit has to cost some precision, however small. The labs all say quality is unchanged and publish benchmarks to show it, and on benchmarks it probably is. Whether a careful writer can feel the difference over a long piece is a fair question that nobody has really tested in public. Artificial Lawyer raised a sharper one for law firms, where a client who has banned AI on their matters might now be able to check, and where a document passed between firms could end up with several different labs' marks layered inside it.

The company that isn't marking at home

Which brings us back to Andrew Curran's line, and to OpenAI.

OpenAI built a text watermark years before anyone else shipped one. In the summer of 2024 the Wall Street Journal reported that the company had had a working method for about a year and hadn't released it, partly because a survey found nearly a third of loyal ChatGPT users said they'd use it less if it were watermarked. When OpenAI finally launched textGrain in October, it made it the default only in the EU, for ChatGPT and Codex. API customers anywhere can switch it on, but it ships off. OpenAI's explanation was that the regional approach "gives us room to learn from real-world use and feedback."

That's a defensible position, and you can make a decent case that it's the more honest one, given everything the company itself says watermarks can't do. But it does leave a strange map. Claude marks its text everywhere because Anthropic couldn't find a clean way to do it only in Europe. Gemini already marks its text, and Google quietly confirmed in August that this covers the API and AI Studio too, after a support answer had first said the opposite. Grok's API will mark by default from 7 November, worldwide, even though xAI never signed the code. And the most used chatbot in the world marks its text in Europe and leaves it unmarked for most of its users in America.

If I had to bet, I'd say that map doesn't last. Once three of the four biggest labs are marking globally because it's simpler than maintaining a border inside their models, the pressure moves to the exception. And once American schools, publishers and platforms realise that a passage from Claude can be checked and a passage from ChatGPT can't, they'll start asking why.

What I'd actually do

If you build on these APIs, the practical part is short. Read your provider's page this week, because the details differ. On Grok, start logging the `x-watermark` header now, so you know what's marked before it flips on, and don't opt out unless a lawyer has told you an exception really applies, because the console is going to ask you to say so on the record. On Claude's two older 4.6 models, know that changing temperature or top_p takes the mark off, and decide on purpose whether that's what you want. If you're relying on the provider's mark for your own compliance in Europe, keep the defaults.

If you write with these tools, the practical part is shorter still. Pete Sena's advice, "write like someone might check," is right, but not for the reason the post implied. Most of the time nobody will check, and when someone does, the result will be a probability from one company's detector and not a verdict on you. The better reason is the old one. If a client, an editor or a reader would mind that a model wrote the thing, tell them, or write it yourself. Keep your own drafts. The mark only ever says the model touched the words. It can't say what you put into them.

And if you're just reading the internet, the useful thing to take from this week is a sense of proportion. From tomorrow there'll be more machine-marked text in the world than there was on Friday, and the people best placed to spot industrial-scale abuse of these systems will have slightly better tools to do it. That's genuinely good. But a missing mark won't mean a person wrote something, and a found one won't tell you who did, or why, or whether it's true. We're still going to have to read things carefully.