AI Text Watermarking Is Free And Good

Zvi Mowshowitz, writing on LessWrong, is pushing back on the backlash over AI text watermarking, arguing the technique is free, effective, and good. The method, developed by Scott Aaronson while he was at OpenAI, works by using a secret key to subtly bias the randomness already present in how language models choose their next token — a detector can later check for that pattern. Google has been implementing it since 2024, including for Gemini 3.7 Flash, and confirmed in a twenty-million-sample test that users noticed no difference. Anthropic quietly announced a week ago it would add watermarking to Claude outputs to comply with the European Union's Code of Practice, and OpenAI intends to follow. Mowshowitz says the entire practical effect is an API that tells you if a piece of writing came from Claude — and that much of the anger stems from what he calls Anthropic derangement syndrome, rather than genuine technical concerns.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton