crossposter.mochi.is

I care about whether he's right about his core claim: that quality is being sacrificed for the sake of a watermark that OpenAI found was "trivial" to circumvent. I think he is. Have written more here:

We're just going round in circles. I quoted the relevant part in paper you can look up, notably the "diversity/detectability trade-off". They themselves connect it to quality loss in one example: "could lower the overall quality of a system which generates many responses then selects the best one".

Define "good". If diversity is good, and contributes to overall quality, reducing it with watermarking is bad, and reduces the overall quality. I understand people saying the loss isn't significant, but don't understand those saying there is no loss whatsoever compared to not watermarking.

Saying "response is one it could have given you anyway" is similar to saying a loaded die can also throw the same number as a fair die. Sure, but no one would accept that it's the same thing as playing the game with a fair die.

Don't think that's right.
1. They're not being constantly re-weighted, otherwise the authors of the paper would not write responses "are more likely to be similar to each other if they are watermarked"
2. Ok, let's say "less fare", analogy still holds, creating a "diversity/detectability trade-off"

Fair to argue about how John presented the LLM selection process, but the main thrust of his piece is that LLM quality is being sacrificed "for the purpose of embedding hidden clues within the text to suggest its provenance." He's right about that.

You're talking about changing a "fair" random die to a loaded die. To an outside observer, the results seem fair, sure. And you can say the die doesn't advantage any particular player, because it's loaded randomly. But if caught, can you really say a loaded didn't skew the results of the game?

Kinda agree, feels like focus has shifted to coding at the expense of writing. Probably where all the money is. But not that long ago, AI labs did seem to care about the quality of writing their models produced. Just look at the GPT 4.5 and 5 announcements compared to GPT 5.4. Search for 'writing'.

It's not as random as before. There is a "diversity/detectability trade-off" and that "the similarity between the responses is greater for the watermarked responses than the unwatermarked" according to the paper Anthropic linked to. media.springernature.com/original/spr...

Text watermarking reduces the diversity overall, which is a reduction in quality. The authors of the paper Antrhopic linked concede there is a "diversity/detectability trade-off" and that "the similarity between the responses is greater for the watermarked responses than the unwatermarked".

And on quality loss, which John appears to have gotten a lot of pushback on, the authors concede there is a "diversity/detectability trade-off" and that "the similarity between the responses is greater for the watermarked responses than the unwatermarked". media.springernature.com/original/spr...

Editing is not exempt, it depends what you change. And Translation was not exempt in May of this year, last month it became exempt, it seems. And it's not a moot point, because whether it's exempt or not in the EU act, Anthropic is watermarking everything. But can you blame them reading it all?

I see two problems with it:

1. There is huge value in using AI systems for many types of work, e.g. translation, editing. That work will now be tarnished as "AI slop" or "disinformation".
2. Watermarking impacts the quality of the AI output. The authors of the paper Anthropic linked admit as much.

I disagree. I think it's hurting ordinary users. There's so many ways to use AI that has nothing to do with generating and distributing fabricated content. Do you think "disinformation networks" are going to be thwarted by a technology that's very easy to circumvent? And I say this as an EU citizen.

I think OpenAI's position on this in 2024 was the right one. It's a shame they gave in too:

Watermarking is, according to 2024 OpenAI:

- "trivial to circumvention by bad actors"
- "could stigmatize use of AI as a useful writing tool for non-native English speakers"

Funnily enough, the EU had an exception for "translations" being watermarked, but removed it in their 2026 update to the act. Anthropic has gone beyond what the act required, but I also don't blame them - read the EU act and their examples and exceptions, and you'll see what a mess it is.