Anthropic details Claude's watermarking approach amidst EU AI Act compliance and user backlash

Anthropic released a blog post on Friday detailing the watermarking of text generated by its chatbot Claude. The move, announced earlier this week to comply with the EU AI Act's Transparency Code, sparked debate among users, with some on Reddit calling it a conspiracy and others on X canceling subscriptions. The post explains that Claude creates undetectable patterns in low-stakes word choices, like between 'overcast' and 'grey', which can be detected with a key. Anthropic will use the SynthID-Text approach from Google DeepMind and plans to release a watermark detection API. It also distinguishes watermarking from AI detection methods like Pangram's, which look for stylistic 'tells'. The company admitted that a complete rewrite eliminating every watermark word would remove the watermark, but light editing probably won't. For proofread or edited text, the watermark's presence depends on the length and extent of editing. Code will have less watermarking because the model must produce functional code, though comments may contain watermarks. Finally, Anthropic noted that other major model developers have signed the same Code of Practice and will implement their own watermarks.