How do AI text watermarks work?

So, Anthropic's finally shed some light on how it's planning to use invisible watermarks on text generated by Claude, all in an effort to comply with those stringent AI transparency rules in Europe. It turns out that Claude's text marking system is actually based on the SynthID-Text approach - an open-source watermarking tech developed by Google DeepMind. This clever technology creates detectable patterns using wording probabilities, which allows the generated text to be identified as artificially generated or manipulated.

The whole point of introducing this watermarking feature is to meet Anthropic's obligations under the European Union's AI Act, which demands that synthetic audio, image, video, and text include machine-readable marks that enable the content to be detected as artificially generated or manipulated. Thankfully, Anthropic assures us that these text watermarks won't make Claude more expensive for users, nor will they have any practical impact on the quality or content of Claude's outputs - that's a relief!

Now, here's how the watermarking process actually works: it uses low-stakes choices in generated text to leave a pattern in Claude's responses. For instance, when generating a sentence, the model might choose between two words that are equally likely to follow the previous words - in cases like this, the choice is settled by a random number. The watermarking tech uses these low-stakes choices to leave a pattern in the text, which is undetectable to the reader but detectable to anyone who has a key that encodes it.

As Anthropic points out, the EU's AI transparency requirements are going to impact other major AI developers too, so Claude won't be the only model introducing text watermarks. Other big players like Google and OpenAI will also be subject to the law's requirements, and may well introduce similar watermarking technologies to their AI models. So, users of AI models can expect to see more widespread adoption of text watermarks in the future, as companies scramble to comply with the EU's AI transparency rules.

Source: The Verge