AI-Generated Text Watermarking Schemes Face Scrutiny Over Quality and Transparency
Discussions surrounding watermarking schemes for AI-generated text, such as those implemented by Anthropic/Claude and Google/Gemini, are raising concerns about their impact on prose quality and transparency. Critics argue that these schemes, which embed hidden signals in AI-generated content, cannot work without degrading the quality of the text, even if subtly. The process involves applying predictable randomness to word choices, which some believe inherently makes the output slightly worse than text generated without such constraints. A significant objection is the secretive nature of these watermarks, which are dependent on proprietary information held by LLM providers, making them unverifiable by users. This lack of transparency is deemed unacceptable by some, who advocate for openly verifiable watermarking if such schemes are to be used. 
