On Friday, Anthropic released a blog post aiming to address fundamental inquiries concerning the method it intends to apply for watermarking the text produced by its chatbot, Claude. For example: How does the watermarking process function? Is it possible to conceal it through editing? What impact does this have on the code? Users of Claude have been discussing the introduction of watermarking due to the company’s compliance with the EU AI Act’s Transparency Code. This code necessitates the use of systems that enable the identification of AI-generated content. Debates on Reddit have ranged from accusations of a conspiracy against innocent Claude users to suggestions that opposition stems from a desire to deceive people. According to Business Insider, numerous users of X have reportedly cancelled their Claude subscriptions. In a recent post, Anthropic begins by providing a comprehensive explanation of the watermarking technique. It states that when faced with ‘low-stakes decisions’, such as selecting between ‘overcast’ and ‘grey’ to depict the weather, Claude can generate an undetectable pattern in its replies. However, this pattern can be identifiable to anyone possessing a key that deciphers it. The company stated that watermarking has no effect on the quality of Claude’s output. “A watermarked reply appears identical to an unwatermarked one to a reader.”.
