What's this about?
Anthropic responded to days of criticism with a detailed FAQ and technical explainer. Claude uses SynthID-Text, the method DeepMind published in *Nature* in 2024, which traces back to a 2022 proposal by Scott Aaronson. Anthropic's guarantees: nothing is added to the text, there are no hidden characters, it requires no extra tokens and costs nothing more, it carries no information identifying a person or chat, and — the key claim — the model never picks a word it wouldn't otherwise have picked. It's also not Claude-specific: since August 2, the EU requires providers serving its market to mark AI-generated content, and other major labs signed the same Code of Practice. The response was not calm. Within days a MIT-licensed watermark stripper reached 10,000 GitHub stars, targeting Claude, SynthID-Text and OpenAI marks plus C2PA and EXIF in images and PDFs. John Gruber called the feature "a perversion of writing," and *Nature* reported researchers' scepticism.
What exactly does "changing the source of randomness" mean?
Anthropic's own example is the clearest. Suppose the model has written "it's cold today, and…" — the next word might be "overcast" or "grey," and to a reader these are equivalent. Normally the model uses a random number to pick between such near-ties. With watermarking, the choice is still random, but the random number is now derived from a secret key combined with the preceding words. The text reads identically. But afterwards, someone holding the key can check the sequence of choices and compute the probability that Claude wrote it. This is why the "no quality loss" claim is defensible: the watermark only steers choices the model was already indifferent between.
What's the community saying?
- Gruber's objection, and the rebuttal that landed on it. He argued he wants an LLM choosing "the very best, most precise words at every decision point." The most-upvoted reply pointed out this already isn't how they work — models sample randomly from a probability distribution at every token, so the watermark changes the source of the randomness, not the presence of it.
- The proofreading problem is the real one. The sharpest practical concern: nobody can safely use Claude to polish their own prose anymore, because the whole piece may then be flagged as Claude-generated. Several commenters believe that outcome is intended.
- The terms-of-service clause alarmed people. The requirement that providers forbid users from removing the watermark raised ownership questions — if you rewrite a section, or edit some lines of code containing the mark, what exactly have you violated?
- Does it even work outside prose? Repeated technical questioning about structured or constrained output: if you ask for "true or false," or for JSON matching a schema, there is no room for the near-tie word choices the scheme depends on.
- The stripper repo settled the practical question. Ten thousand stars in days on a tool that removes these marks says more about enforceability than either side's argument — and it targets every major provider's scheme at once.