Anthropic Adds EU-Mandated Watermark to Claude AI Model

Anthropic announced that its Claude large-language model will embed a watermark in all generated prose to comply with a new European Union regulation that obliges AI providers to label AI-generated text by December. The company says the watermark will be applied at the “granular, random level” where the model selects words, and that it will remain invisible to most readers while being detectable by Anthropic and parties possessing a decoding key.

The Guardian

The EU rule, which applies to any AI service operating in the bloc, aims to curb undisclosed AI-generated content and mitigate “model collapse” – a degradation of model performance caused by training on AI-written text. Anthropic’s implementation is intended to preserve model integrity and provide a tool against disinformation.

The Guardian

Anthropic’s release emphasizes that the watermark will not alter the underlying randomness of Claude’s word-selection process. Computer-science professor Steven Murdoch of University College London echoed this view, noting that the change makes the random number generators “statistically predictable, but still random.” “There’s going to be no noticeable difference. There’s the same random number generators there – it just used to be completely random, and now it’s statistically predictable, but still random.” — Steven Murdoch, a professor of computer science at University College London He added that the watermark could help prevent the proliferation of AI-written material that might otherwise degrade future models. “There’s already randomness involved in any of these large language models,” — Steven Murdoch, a professor of computer science at University College London

The Guardian

If the watermark functions as described, students, lawyers, and academics may find it harder to pass off AI-generated drafts as original work. At the same time, the added traceability could reduce the risk of “model collapse” by limiting the feedback loop of AI-on-AI training data. The ultimate effect on Claude’s prose quality remains an open question, pending real-world deployment later this year.

The Guardian