Anthropic has clarified the functioning of an invisible watermark intended for texts produced by Claude. This technology utilizes SynthID-Text, developed by Google DeepMind, and will be integrated during the response generation process.
The main objective of this initiative is to establish a verification method for content generated by the platform itself, aligning with the guidelines of the European Union's Artificial Intelligence Act. This European regulation came into effect on August 2nd, requiring that content created by AI tools be traceable, suggesting the adoption of similar technologies in other models.
Although SynthID is already known for applying watermarks to media such as video, images, and audio, introducing subtle modifications imperceptible to the human eye, SynthID-Text operates by altering the pattern of words selected by the model without compromising text quality.
How does the watermark work?
In an official statement, Anthropic described that the identification is built sequentially based on the words generated by the assistant. Language models construct text by calculating the statistical probability of each word, considering the previous context, and this is where the watermark comes into play.
To illustrate, in a sentence like 'The weather today was cold and…', the AI would assess that the improbable continuation would be 'sweetened', but it could be 'cloudy' or 'grey'. When there are multiple options for synonyms, the model makes a random decision about which word to use. With the implementation of the watermark, an encryption key guides this selection.
Subsequently, Claude will analyze whether the found pattern is consistent with a text generated by itself. The company emphasizes that this does not imply that Claude will start inserting inappropriate terms into the middle of the text, as internal tests and studies of SynthID-Text conducted by Google DeepMind indicated that the technique does not affect the final content.
Where might detection fail?
Due to the way the mark is applied to the writing pattern, Anthropic points out that simple modifications to the text can impair the analysis, and substantial changes to the content structure tend to eliminate the marking.
Furthermore, there are other circumstances where detection may fail, such as when the excerpts are very brief, when the answer is based on precise factual data, in mathematical operations, in programming code, or in texts written by humans but subsequently revised by the assistant.
Finally, Anthropic announced that it will soon make a detection API available, allowing developers, institutions, and platforms to verify the probable origin of a text generated by Claude.