What's this about?
Anthropic has shipped a public tool for checking whether a file was produced with Claude, using the text watermark it announced in August.
How does a text watermark work?
Not by inserting invisible characters — that is the common misunderstanding and it would be defeated by pasting as plain text. The technique biases the model's token selection according to a secret key: at each step, some tokens are made slightly more likely than they otherwise would have been. Any individual choice is unremarkable; across a long enough passage the statistical fingerprint is detectable by whoever holds the key. It survives light editing and format changes, and it degrades as the text is rewritten.
What's the community saying?
HN 152 points, 112 comments
- The top comment reframes the question well: a watermark answers "did this model touch the file?" It does not answer "did a person read it and stand behind it?" — and the second is the one that matters.
- A confident dismissal ("defeated by a text editor and Ctrl+Shift+V, Unicode characters are nothing new") gets corrected: that is not how this watermark works.
- The credible objection that survives: pass the output through another LLM and the signal likely goes. It may catch students; it will not catch spam operations, which will add an anti-watermark pass the moment platforms start checking.
- A partial defence: bypassing it without distorting the original message may be harder than it sounds.