AI Text Watermarking Is Free And Good
Zvi (Don't Worry About the Vase) TheZvi ● Covered by 2 sources
Anthropic is adding text watermarking to Claude to meet the EU code. It’s meant to be free, invisible, and easy to check, which is why people are arguing about it.
Based on reporting by Zvi (Don't Worry About the Vase), TheZvi — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Scott Aaronson, working with Hendrik Kirchner at OpenAI, helped solve a problem that had looked awkward for AI text: how to mark generated words without changing them in any noticeable way. The trick leans on something simple. LLM outputs are already random, because the model picks probabilities for next tokens and then samples from them. Watermarking swaps in a private pseudo-random source tied to a secret key, then checks later whether the text lines up with that source better than it would with another one.
The appeal is obvious. The method is designed to leave outputs effectively unchanged for users, while letting an API detect whether a piece of text came from a model. The source says humans can’t tell the difference, that the marginal cost is very close to zero, and that the mark can be washed out by rewriting in your own words. In other words: no dramatic visible scar, just a quiet signal hidden in the token choices.
That matters now because the European Union Code of Practice requires future AI models to use watermarks, and the major Western labs signed it. Google has already implemented this, including for Gemini 3.7 Flash, and has been rolling the feature out since 2024. The source also says Google tested it with 20 million cases and found no difference in user feedback. Anthropic said a week ago that it was rolling out watermarking to comply with the same code, and that because it doesn’t want to separate traffic sources, the marginal cost is zero.
The reaction has been louder than the change. A lot of the anger, the source argues, is really about Anthropic itself, or about the idea of any alteration to model output at all. Some people dislike the word “watermark.” Some just don’t want AI text to be identifiable. But the core claim here is plain enough: if the math works and the cost is basically nothing, then watermarking is a useful thing to do. The real controversy is less about the technique than about who gets to feel anonymous.
My take — AI-written commentary, not fact-checked reporting
This is one of those rare AI moves that sounds boring because it is mostly just competent. If a watermark is free, invisible, and already required by the rules, the outrage reads less like principle and more like people annoyed that their disguise got slightly worse. The industry has spent years pretending every hidden mechanism is a plot; sometimes it’s just plumbing with better manners.
Read more about this at: Zvi (Don't Worry About the Vase)