Watermarking of AI test under Claude

[Robert Brown (RLBrown)]

Hello SVMUG members and friends,

The EU AI Act, Article 50,  and the Code of Practice on Transparency of AI-Generated Content require that AI-generated material be detectable. Anthropic is complying with this new law by using SynthID Text, a method developed by Google DeepMind in 2024. 

The method relies on choosing synonyms for words in generated text. Normally, if multiple equally valid word choices are available for a given position in the text, a pseudo-random choice might be made. For SynthID, the choice is made by applying a secret cryptographic key to the context preceding the word in question. This slightly biases the word choice to a mathematical pattern. 

The text does not use hidden Unicode, zero-width spaces, or awkward phrasing. The text reads naturally. However, it is only reliable for reasonably long pieces of text. Furthermore, it can not be applied to text that is inherently subject to construction rules, for example, computer code or mathematical formulae. 

Anthropic provides a dedicated detection endpoint that analyzes submitted text against the watermark key to return a likelihood score indicating whether Claude generated the text. There are no local tools, such as an application you could run on macOS, to detect the watermark. However, it is suggested that a Python or terminal script could be created to automatically submit text to the endpoint for analysis. 

Another failing of this watermarking method is that light editing or rephrasing of the generated text will destroy the watermark. 

Claude, Gemini, and Siri assisted in this report. But you can’t tell. Who says “formulae” instead of “formulas"?

Robert Lawson Brown mail