Anthropic’s LLM watermarking

· Scott Aaronson · Aug. 22, 2026, 10:24 p.m.
Summary
Scott Aaronson discusses Anthropic's announcement of watermarking outputs from their AI model, Claude, referencing a previous proposal he made while at OpenAI regarding LLM watermarking. The watermarking utilizes a scheme based on Google's SynthID, which builds on Aaronson's Gumbel Softmax method from 2022, marking a significant step in the field of large language models.