Anthropic Is Embedding Invisible Watermarks in Claude's Text

Anthropic has shared the details of how it plans to place invisible watermarks on text created by its Claude model. On Friday the company said that its labeling system is “a version of the SynthID-Text approach,” an open technology developed by Google's DeepMind division. It produces patterns that are imperceptible to the reader but detectable, by shaping the probabilities of word choices.

Text watermarks, along with support for the C2PA standard for images processed by Claude, are being introduced to fulfill Anthropic's obligations under the European AI Act. The law requires that synthetic audio, images, video and text carry machine-readable markers that make it possible to identify content generated or altered by AI. According to Anthropic, the watermarks will not make Claude more expensive for users, nor will they have “any practical impact on the quality or content of responses.”

How it works

Anthropic explains the mechanism with an example. Take the phrase “The weather today was cold and…” The next word is unlikely to be “sugary,” but “overcast” or “gray” are entirely plausible. The reader usually doesn't care which of the fitting options the model picks — the meaning of the sentence hardly changes. In such cases the choice is settled by a random number.

It is precisely these “low-stakes” forks, which occur many times throughout a text, that are used to imprint the pattern. It is invisible when reading, but detectable by anyone who has the key that encodes the pattern. With labeling enabled, the choice is still made at random, but the source of the randomness changes: instead of an arbitrary random number generator, the model relies on the key and a few preceding words to determine the next word.

Not just Claude

As Anthropic notes, the EU's transparency requirements apply to other major AI developers as well, so Claude will not be the only model with text watermarks. Google's Gemini chatbot has supported SynthID-Text since 2024. OpenAI has not yet disclosed any plans to label text for ChatGPT in its AI Act compliance roadmap, but the company falls under the law too.

Source: The Verge