Your AI Now Signs Its Work. Invisibly.

Back in early August I wrote about the EU’s new AI transparency rules and said the AI companies would have to start embedding invisible, machine-readable markers in what they generate. That was the provider half of the law, and I said it was already happening under the hood.

It’s now happening on top of the hood. Anthropic confirmed on August 11 that Claude models launched on or after August 2 embed an invisible watermark directly into generated text, and attach signed provenance metadata to generated files like images and SVGs. It applies worldwide, not just in Europe, and there’s currently no opt-out. Models released before August 2 are getting it during a transition period over the coming months.

If you publish content, this is worth ten minutes of your attention. It’s also worth cutting through some noise, including some I contributed myself.

What it actually is

The watermark is not a phrase, a character, or a hidden symbol. It’s a statistical bias in word selection made while the text is being generated: Claude nudges its own choices in a pattern that a detector holding the right key can spot across enough text. Anthropic describes it as a version of the same approach Google DeepMind published for text watermarking, and says internal testing showed no meaningful quality difference.

Two consequences worth understanding.

The mark travels with copied text. Paste it into a document, an email, or your website, and it goes along. Unlike file metadata, it doesn’t get stripped by resaving or format conversion.

And you cannot see it. This is the part I want to correct, because I said otherwise on a recent call and I was partially wrong. Those puffy, consultant-flavored word choices you’ve noticed lately, the endeavored and the facilitate and the unnecessary adjectives, are not necessarily the watermark. That’s the ordinary house style of AI writing, the same tells that AI detectors have always looked for. Deleting those words improves your writing. It does not always remove the watermark.

What it does not do

Third-party AI detectors cannot read it. They don’t have the key. They’re still doing what they always did, which is guessing from style. So the watermark does not make you more likely to get flagged by the detector your client or your platform runs.

It also isn’t proof of much on its own, and Anthropic says so plainly. Unmarked text doesn’t mean a human wrote it. Marked text doesn’t mean a human didn’t. A translation carries a full watermark because every word was chosen by the model, even though the ideas are entirely yours. And when the model lightly proofreads something you wrote, most of the words are still yours, so there’s little signal to detect.

Heavy editing weakens it. Short passages may not carry enough signal to detect at all.

What I’d actually do about it

Nothing dramatic, and here’s why.

If you’re publishing AI-assisted content under your name without saying so, this news should make you uncomfortable, and the discomfort is the useful part. The fix isn’t stripping watermarks; it’s the thing the EU rules already pointed at and that good practice always suggested: a human reviews the work, a named person takes editorial responsibility, and if it matters to your audience, you say so. That’s what the footer of this site does. I write with AI assistance (this post is partly built from AI pulling content from my recent talk on this subject and adding it to this article), I edit everything, my name is on it, and the disclosure is public. A watermark on top of an already-honest process changes nothing for me.

If your work involves translating or editing other people’s writing, the calculus is different and genuinely unresolved. Your client’s document could come back carrying a mark suggesting authorship that isn’t yours or theirs. That’s a real problem, the vendors know it, and it’s the loudest complaint in the response so far. Watch it, and don’t let anyone tell you a watermark settles an authorship dispute.

And if you’re relying on AI-detection scores in either direction, hiring, grading, vetting freelancers, this is your reminder that those tools were never reliable and this doesn’t fix them. A detection API is coming from the vendor for its own mark. Until then, and honestly after, judge the work.

The broader signal matters more than the mechanics. The AI companies are now marking output at the source, worldwide, because a European law told them to. Whatever you think of the rule, the era of quietly passing off machine work as your own is closing, not because of the watermark, but because everyone is building for a world where provenance is expected. Getting honest early costs you nothing. Getting caught late costs plenty. The rules that started this are here: https://tablelandpartners.com/eu-ai-content-rules-business/