Voices
Zvi Mowshowitz: text watermarking is free because LLM outputs were already random, and Google proved it at n=20 million
Zvi's August 21 post attacks the intuition that watermarking degrades output. Because each token is already sampled from a cloud of near-equivalent options, swapping the sampler for a keyed pseudorandom source encodes a detectable signal at effectively zero marginal cost, and Google confirmed no difference in user feedback across a 20-million-sample test after two years of shipping it in Gemini. He notes Anthropic announced its rollout quietly to comply with the EU Code of Practice, that OpenAI intends to follow but looks likely to miss the deadline, and that paraphrasing removes the mark in proportion to how many of the model's word choices you keep.
↳ Follow the thread