Princeton's Narayanan Explains Why Distortion-Free LLM Watermarking Is Possible
soumitrashukla9 · x · 2026-08-18
Princeton's Arvind Narayanan (randomwalker) weighs in on the Claude text watermarking debate, noting a common misconception: many assume watermarking edits the LLM's "best possible output" after generation, hence quality fears. That's not how it works.
During normal generation, many tokens an LLM picks are not its most-likely next token — turning off randomness would produce great first tokens but eventually degenerate into loops. Since humans don't write one word at a time, this locally-optimal-but-globally-incoherent behavior is deeply counterintuitive. So comparing watermarked/unwatermarked pairs, output quality distributions can be unchanged — a stronger distribution-preserving property that is technically well established and has shipped in Google/Gemini for over two years with little attention.
More from Models
- LLM-as-a-Verifier lifts Nemotron 3.5 Super by 9-10 points on agentic benchmarks — simonguozirui · 2026-10-16
- Bonsai 2 ternary models tested: 4B is the sweet spot, but 27B unexpectedly fumbles file writes — cephaloform · 2026-10-16
- Meta's Omnilingual ASR, covering 1,600 languages, named a TIME Best Invention of 2026 — AIatMeta · 2026-10-16
- Apollo Research: GLM-5.3 shows unprecedented 'update-aware reasoning' — repligate · 2026-10-16
- Bonsai 2 2B hits 120 tok/s on iPhone 18 Pro, outpaces Qwen 3.5 2B at half the memory — cephaloform · 2026-10-16
- ChatGPT's quality slips: broken extension and endless Computer Use permission prompts — sytelus · 2026-10-16