Anthropic details Claude’s text watermark and plans a detection API
Anthropic published a post on Friday setting out how Claude’s text watermark actually works, filling in the mechanism it left out of its original watermarking announcement four days earlier. The company says Claude’s watermark is a version of the SynthID-Text approach that Google DeepMind published in a Nature paper in 2024. It also said a watermark detection API is coming, though the details aren’t settled yet.
The mechanism sits in word choice rather than in hidden characters. Anthropic’s example is the sentence “The weather today was cold and…”, where the next word is unlikely to be “sugary” but quite likely to be “overcast” or “grey”. Normally that choice is settled by a random number. Watermarking swaps the source of that randomness for the key plus a few preceding words, as The Verge reported.
That pattern is undetectable to the reader, but is detectable to anyone who has a key that encodes it.
Anthropic, via The Verge
That matters because the driver here is regulatory, not product. Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, one of around 190 signatories. The EU AI Act requires synthetic audio, image, video and text to carry machine-readable marks that flag the content as artificially generated. Claude-processed images get C2PA support under the same obligation.
Anthropic’s post also answers the questions users kept asking about cost, speed and privacy. Here’s what the company said on each.
| Question | Anthropic’s answer |
|---|---|
| Does it cost more? | Watermarking “produces no extra tokens”, so the model is “the same price to serve and use” |
| Is it slower? | “Negligible impact on the speed of models” |
| Does quality drop? | Testing showed no impact on quality |
| Can it identify me? | Nothing in the watermark or its key recovers information about the user, their organization, or their chats |
| How do I check text? | A detection API is coming, with implementation details still being worked out |
The limits are where it gets interesting. Anthropic says light editing probably won’t remove the watermark completely, while a complete rewrite that replaces every word does. As the company put it, at that point “it’s arguable whether the text can any longer be described as AI-generated”. Detection also works poorly on short samples, because there are fewer word choices to hide a pattern in.
Code carries less of a watermark than prose, because working code doesn’t leave the model many equally valid options. Anthropic says the marking can still land in places where the choice is arbitrary, such as comments, with what it calls a negligible effect on the code itself. Google DeepMind’s own write-up flags the same weak spots, noting that confidence scores “can be greatly reduced when an AI-generated text is thoroughly rewritten or translated to another language“. Factual passages give the system fewer openings too.
None of that is the same thing as an AI detector. Anthropic drew the line explicitly, saying tools from companies like Pangram look for writing tells, and that spotting those patterns is fundamentally different from checking for a watermark, TechCrunch reported. If you’ve followed why most detectors fail, the distinction is the whole argument: one needs a secret key, the other guesses from style.
The critics aren’t arguing about the plumbing. Daring Fireball called the scheme an adulteration of writing, on the grounds that swapping one word for a statistically convenient near-synonym isn’t free.
No two synonyms carry the exact same meaning.
Daring Fireball
His example is “He leaped at the chance” against “He jumped at the opportunity”, two sentences carrying the same sentiment without being the same. Anthropic’s counter is that a watermarked response is indistinguishable from an unwatermarked one to a reader, which is a claim about perception rather than precision. Those two positions don’t actually contradict each other, and that’s why the argument hasn’t resolved.
The backlash has been noisy. TechCrunch reported that Reddit users have been trading accusations over the move, and that Business Insider reported “dozens” of users on X claiming to cancel their subscriptions. Anthropic’s response is that Claude won’t be alone, because other major developers signed the same Code of Practice and are building their own marks.
Gemini has supported SynthID Text since 2024, so Google’s already through this. OpenAI hasn’t detailed any text watermarking plans for ChatGPT in its AI Act compliance roadmap, though the law reaches it all the same. Watch the detection API: who gets access to it, and whether anyone outside Anthropic can run a check, decides whether this is a compliance artefact or something a reader can use. Google’s decision to make Gemini’s visible watermark optional is a reminder that these policies move.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
