Anthropic clarifies doubts about the watermark that Claude will apply to his texts

The Anthropic Artificial Intelligence (AI) laboratory has given more details to clarify some issues that were not completely clear about the watermark that will be applied worldwide in the texts generated by Claude in the application and web, API, Claude Code, Claude Cowork and Claude Tag.

Last week, the company revealed that will apply watermarks to the texts generated by Claude after the signing of the Code of Good Practice on the Transparency of Content Generated by AI that is part of the AI ​​Law of the European Union.

Within days, Claude users went from criticizing the measure to stopping paying for Anthropic’s AI subscription as a result, according to the specialized media Business Insider.

Anthropic has given more details for clarify certain aspects in which there seems to be confusion. The first is that the algorithm used is a version of SynthID-Text (developed by Google DeepMind), which is based on the low-stakes choices between equally valid words and it only changes the source of randomness with which the model chooses, using a key.

That is, it uses a mathematical key and the previous words to decide the next word to embed the watermark. In this way, Claude creates in his answers a pattern that is undetectable to the reader, but is recognizable to the person who has the key that encodes it.

The AI ​​laboratory has pointed out that the embedding of the watermark does not entail latency or an extra cost in tokens or in the API, and does not store any type of personal data of the chat user or organization.

Furthermore, he points out again that the brand is not a resounding yes or nobut a probability. This means that if it detects the pattern, it will say “probably passed by Claude”, although it doesn’t work the other way around, since if it doesn’t find a mark, it doesn’t mean a person wrote it.

In this sense, Claude includes his own key and It differs from ChatGPT or Geminiso there is no universal detector. And for the detection to work better, the more words, the more reliable the brand will be.

Likewise, in short texts it barely offers a signal because the pattern cannot be built on so few words. In the case of code or fixeseverything changes, since the brand cannot be embedded as there is no free choice between words.

Finally, Anthropic has confirmed that it will soon launch a detection API and that every translation that is made does carry a mark for something quite simple to understand: each word is chosen by Claude.