Post

Alex Hern
@hern.bsky.social

If you use Claude right now, then it doesn’t pick the “best” word for a sentence. It creates a distribution of possible words, assesses the probability of any given word fitting the slot, and randomly picks between them based on that probability.

John Gruber
@gruber.foo

I addressed this in a follow-up post yesterday. The per-token randomness in existing models is entirely in the interest of generating semantically better output. The fingerprinting scheme are not to improve output but to embed a signal in the output: daringfireball.net/2026/08/follow…

daringfireball.netFollow-Up Thoughts on Watermarking Schemes for AI-Generated TextI want the answers that I read to be cogent, lucid, accurate, blessedly terse — and ideally to strike a consistent tone that is pleasant to my reading ear. The genie is not going back in the bottle.

0 reposts2 likes

Alex Hern
@hern.bsky.social

Thanks John. It’s a good follow up, and clear we now just disagree on the principle. For me: When a system produces a different answer every time you pull the trigger, I don’t see a big problem with constraining it to a subset of the answers it would already have produced

John Gruber
@gruber.foo

By my reading, Google's SynthID paper makes clear that if they limit themselves to applying the watermark to choices that are truly equivalent (per the model's PLM function), the watermark can't reliably be detected.

John Gruber
@gruber.foo

But the choices aren't from unranked lists. They're from ranked lists, and the only way to leave a signal strong enough to detect is to promote lower-ranked choices. The more subtle the promotion of lower-ranked choices, the weaker the fingerprint. Stronger signal necessarily means worse choices.