The fight over AI text watermarks isn't about detection — it's about who controls the randomness that makes language work.
The Summary
- John Gruber responds to critics of his anti-watermarking stance, clarifying that watermarking schemes inject detectable patterns *on top of* the temperature randomness that makes LLMs produce good prose
- Temperature controls how LLMs select words: 0 = always pick the top choice (robotic), 1 = weighted randomness (natural), >1 = more chaos
- Watermarks add *predictable* randomness for detection purposes, not quality purposes, which Gruber argues inherently degrades output
The Signal
Gruber's follow-up cuts through the noise around Anthropic's watermarking implementation in Claude. The core technical point: LLMs don't pick the "best" word at each step. They use temperature-weighted randomness because deterministic output (temperature 0) produces stilted, repetitive text. The randomness *improves* quality, counterintuitively.
Watermarking schemes layer a second kind of randomness on top. This second layer is predictable if you have the secret key, detectable if you're looking for the pattern. But unlike temperature randomness, it serves detection, not quality. That's the rub.
"The temperature-based randomness is in place to help make the output better. The watermarking schemes are applying predictable randomness for an entirely different purpose."
Watermark defenders claim no quality loss because they don't change temperature settings. They only shift which tokens get selected within the temperature's probability distribution. Gruber's not buying it. If you're biasing token selection toward a detectable pattern instead of toward better prose, you're making a tradeoff. Maybe small. Maybe unnoticeable in most contexts. But real.
This matters beyond academic arguments about prose quality. As AI agents become infrastructure, watermarking becomes policy. California's SB 942 would require watermarks on AI-generated content. The EU AI Act includes disclosure requirements. These laws assume watermarking is a neutral transparency tool.
Key stakes:
- Watermarks create a permanent quality tax on AI output for regulatory compliance
- Detection schemes assume adversarial use cases, penalizing legitimate agent workflows
- Once embedded in law, watermarking schemes become unchangeable even as models improve
The deeper issue: who decides what randomness is "good" randomness. Temperature randomness serves the model's core function. Watermark randomness serves someone else's detection needs. Stack enough "someone else's needs" on top of the base model and you end up with the AI equivalent of a website crushed under cookie banners and tracking scripts.
The Implication
If you're building on LLMs, watch the watermarking policy fight. It determines whether your agents' output quality gets taxed for regulatory legibility. The companies pushing hardest for mandatory watermarks are the ones with the most to gain from detection moats — incumbents who can absorb the quality hit and compliance overhead.
For now, temperature is still yours to control. Use it. Understand what it does and why it matters. Because the next battle is over who gets to inject what else into your model's decision tree.