Anthropic’s LLM watermarking

Scott Aaronson, on his blog Shtetl-Optimized, reports that Anthropic has begun watermarking the outputs of its Claude model, using a scheme based on Google's SynthID and a method he proposed back in twenty-two. Watermarking embeds a subtle signal that can later prove a text came from a specific AI, without hurting output quality. The main drawback, he notes, is that the watermark can be removed with extra work, like translating or paraphrasing. Still, Anthropic says anyone will be able to detect the watermark, and OpenAI suggests it may follow suit.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton