The Grammar Check That Outs You as a Machine

Anthropic started embedding invisible watermarks into every piece of text Claude generates. If you've used Claude for a draft, an email, or a chunk of your book this week, that text now carries a signal you can't see, designed to survive copy and paste. This is not a hypothetical rollout: Anthropic confirmed the change in an updated support page, framing it as compliance with European regulations covering AI-generated content.

I sat on this piece for a few days on purpose. Anthropic had only confirmed that it was watermarking Claude's text, not how, and I wanted to write about the mechanism, not just the announcement. On August 14, 2026, Anthropic published a detailed account of how the text watermark actually works, identifying it as a version of the SynthID-Text technique that Google DeepMind published in a peer-reviewed 2024 Nature paper. That's the confirmation I was waiting for, so here's the piece, now with the actual mechanics instead of informed guessing.

The law behind it

The trigger is regulatory, not voluntary. Anthropic, along with several other major AI model providers and around 190 total signatories, signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, which requires AI system providers to use methods of "marking" AI-generated text. Anthropic didn't scope this to Europe. The company says it's applying watermarking globally at launch because it doesn't yet have a durable way to scope it by region. One region's law became the whole world's default, not by legal obligation but by technical convenience. That's the pattern I keep seeing in AI governance broadly: a handful of jurisdictions write the enforceable rules, and the rest of the world inherits the product decisions those rules force, whether their own regulators asked for it or not.

How it actually works

Here's the part I wanted confirmed before writing this. AI models typically generate a word at a time and decide on the next word based on the preceding text. When there's a choice that doesn't matter much to the reader, like whether to say "overcast" or "grey," the model uses a random number to decide which word to generate. Watermarking works by changing the source of that randomness, so the choices still look random to a reader but are decodable to anyone holding the key.

Anthropic has been fairly direct about the limits and guarantees of this system. The difference between watermarked and unwatermarked text won't be distinguishable to readers — nothing is added to the text, and there are no hidden characters. The watermark doesn't identify anything about individual users. There's nothing in the mark, or its key, that would let anyone recover information about the user, their organization, or their chats with Claude. A watermark only helps test whether Claude might have produced or processed the content. It doesn't say anything about ownership or authorship, and doesn't change who owns the output.

That third point is the one I think matters most for this newsletter's audience, and I'll come back to it.

Where it breaks, now with actual numbers instead of speculation

This is the confirmation that matters most for anyone editing Claude's output before publishing, which is most of us. Anthropic says light editing may leave the watermark detectable, while a complete rewrite can remove it. So the earlier reasoning holds up: a grammar and punctuation pass is light editing. It's exactly the kind of touch likely to leave the pattern intact, because it barely touches word choice or sentence structure, which is what the mark rides on.

Anthropic is also building the piece that turns this from a curiosity into something with actual teeth: a watermark detection API is coming, and the company says it's still working out the implementation details. Right now, detection is theoretical for most people. Once it ships, it stops being theoretical.

Why the detection API is the part to actually watch

I'll say this plainly, as someone who spends a lot of my working life thinking about AI governance: the watermark itself is not the consequential part. It's inert until someone can check it. Once verification becomes widely accessible, publishers, companies, and software platforms could start checking Claude-generated material automatically. That's the moment this stops being an abstract transparency feature and starts being an operational filter, in hiring, in publishing, in academic settings, in performance reviews. And the result of that check will need far more careful interpretation than a yes-or-no answer allows: "Claude was probably involved" is not the same thing as "Claude wrote this", and I don't think most platforms building automated checks on top of the API will bother to preserve that distinction. Flat pass-fail signals are cheap to build and easy to misuse. Nuanced ones aren't, and nobody's incentivized to build the nuanced version first.

The part that still gets missed: this punishes light edits, unevenly

If you write something yourself and just ask Claude to fix spelling or punctuation, the confirmed mechanics say that pattern likely survives. You didn't generate the ideas or the sentences with AI. You used it the way you'd use a spellchecker. But a detector, and increasingly, a detector plugged into a hiring pipeline or a publishing CMS, doesn't know the difference between "I wrote this and cleaned it up" and "I asked for this."

For a non-native English writer, that's a real cost, not a technicality. Running a paragraph through Claude to fix grammar and get natural phrasing used to be a quiet, low-stakes way to get your actual thoughts into readable English. Now that same light pass can leave a trace that gets flagged as "AI-written," even though the ideas, structure, and feeling in the piece are entirely yours. The tool that used to lower the barrier to writing in a second language now comes with a detection risk attached to using it for exactly that purpose. And this happens to whoever needs the grammar pass most, which correlates with whoever is writing furthest from their first language, not with whoever is trying to pass off AI writing as their own.

What it doesn't prove, confirmed by Anthropic itself

A watermark only helps test whether Claude might have produced or processed content. It doesn't say anything about ownership or authorship. That's Anthropic's own framing, not a caveat I'm adding. It cuts both ways: a positive detection doesn't prove the piece was AI-authored, only that Claude touched it at some point, to some degree, for some reason. Proofreading, translation, and a grammar pass all leave the same kind of trace as a fully generated draft.

So what's the actual fix?

I don't think the fix is better watermarking. Anthropic has actually built a fairly careful, low-impact version of this already. I think the fix is in how the signal gets interpreted downstream, once the detection API exists and other people start building on top of it.

Regulation and internal policy alike need to separate degree of use from use at all. A light grammar pass and a fully AI-written draft are not the same act and shouldn't carry the same disclosure weight. When the detection API ships, it needs to report something more like a confidence range or likely type of assistance, not a flat yes or no. Platforms, employers, and publications building on this API need to internalize the limitation Anthropic has already stated outright: a watermark doesn't establish authorship, and its absence doesn't establish humanness either. For non-native writers specifically, the fairest fix is cultural, not technical: treating AI-assisted grammar helps the same way we treat spellcheck, which stopped being a disclosure issue decades ago.

Other labs are expected to follow with similar statistical watermarking systems of their own, which means this becomes an industry default fast, not an Anthropic quirk. That makes this the window to get the framing right, before "watermark detected" quietly becomes shorthand for "not real writing," and the people who most need language tools end up paying the highest price for using them.