if I ask it to paint with a shade of red, but it paints with a slightly different shade of red, that is a fucking effect on quality, pardon my watermarking
If you type like Joey using a thesaurus for the first time, it has an effect on quality
The llm never deterministically picks a shade of red. It's a probability distribution over shades of colors, with certain shades of red being more likely than others. Without fingerprinting, it randomly samples from the distribution using a certain pseudorandom RNG. With fingerprinting, it also selects from the distribution using a pseudorandom RNG. My understanding is that the fingerprinted prng is still a strong RNG. Neither output is more correct than the other.
If a certain token is far more likely than any other, it's usually chosen even in the fingerprinted output.
Yes, language is like that, at least the kind of language produced by LLMs. All LLMs produce a probability distribution at each token. If you run the LLM multiple times with the same prompt you will observe it generate different responses. Using the watermarked prng does not change the distribution.
When generating tokens that might be critical to the tone or grammar or correctness, the probability distribution might be 99% on a certain token. In these cases, with or without watermarking, the output will almost always be that same token. E.g., if you ask "please output the exact word watermelon", the LLM will output watermelon with 99%+ probability even with watermark (i.e., the output won't actually be detectable as watermarked).
Language is different; the tone changes when you change any word or even just punctuation.
Language doesn't work that way -- moving the placement of even a comma will affect its tone.
However, language isn't like that: even if you only drop a single piece of punctuation, that can impact the overall meaning of a sentence.
There are many ways of phrasing things that are, for all practical purposes, functionally equivalent.
Putting aside the way you're saying it, your comment has a valid and common misunderstanding of LLMs.
LLMs don't just naturally output a single suggested word (or token) each iteration. Instead, they output a value (roughly, a probability) for every possible word. It seems obvious to simply pick the top (i.e. best) suggestion each time. Then your objection makes sense: watermarking would violate this.
Of course people have tried this! The problem is, in practice this makes the LLM much less "creative" than if you randomly pick one of its suggestions (weighted by the numbers it assigned them). You can artificially increase the value higher-value outputs to reduce the chances of it saying something really odd, and this parameter is called "temperature". A higher temperature allows lower-probability choices (therefore seemingly more creative but perhaps less accurate) and a lower number vice-versa. Either extreme works poorly, and picking a good number is part of optimising an LLM.
It literally re-weights the output tokens from what the LLM would otherwise have chosen. It _has_ to. It can't be positive, because then that's not watermarking, it's a better LLM.
To add nuance, that article does say one of the two versions does reduce text quality (and the other is worse at detectability):
> SynthID-Text can be configured to be non-distortionary (preserving text quality) or distortionary (improving watermark detectability at the cost of text quality).
What if the next token represents a wrong or low-quality answer, but would have only been picked 10% of the time, but now it's picked 20% of the time? Doesn't that obviously decrease the model quality, even though "it might have picked that token anyway"?
It would be picked 10% of the time with watermarking.
The randomness properties of the PRNG will be very similar to other random number generators, it is just chosen to be vulnerable to a particular cryptanalytic attack (that requires a private key known only to anthropic). I think of it like the Dual_EC_DRGB generator rather than a biased coin.
It is absolutely possible that it would not continue to be picked 10% of the time with a given fixed watermark key. The implementation literally labels tokens using a keyed hash and then modifies their scores. The entire point of the watermarking system is to bias certain tokens against others, and - as you would expect - this reportedly results in a reduced response diversity.
What if the token represents a high-quality answer, but would have only been picked 80% of the time, but now it's picked 90% of the time?
I'm not entirely sure (haven't read the original synthID proposal), but I believe that the re-weighing is set to make both your scenarios and mine equally likely, averaging out to net Zero effect on quality.
Unless you’re at 0 temperature, there is no single token it would have chosen. It’s always picking one of multiple randomly according to a probability distribution.
Watermarking just alters the pseudorandom number generator. If "I like turtles" was previously the response to your prompt with probability 100%, it will still be so. This is why watermarking is only effective for long strings of text
It's like the sudden change of a language style and its verbosity didn't happen recently.
To random words you pick and provide a sufficient amount of text to vary with random number without losing its meaning you need a text with high entropy.
Nothing about watermarking would require padding the response length with pseudo-intelligible Claudese. Regular filler would work fine.
Also, it would probably provide higher entropy to write normal human-sounding English instead of reusing a repetitive grab bag of load-bearing phrases. This theory doesn't really make any sense.
No, it basically uses a fixed seed for the random number generator to generate tokens. The generated random sequence is just as random as it would be with a random seed.
I am pretty sure they did A/B testing to show it didn't. I could gave sworn they even released a quiz were the user has to try and guess which answer is watermarked or not and it was impossible to tell.
That's not the case, because LLMs are non-deterministic.
It only alters outputs when the last layer of the neural network give significant weights to multiple tokens, and it would anyway have picked a random answer.
Instead it picks a non-random one, but non-random in such a way that you can't tell without the private key of the watermarking.
This mostly adds randomness these days for branches in syntax that make no difference, and the model has no reason to believe make a difference. Anything that matters, it is much more confident in the last layer of weights on the token to use.
>That's not the case, because LLMs are non-deterministic.
That feels a bit like a lie. At the core, they are deterministic. We found that adding some ability to randomly pick the second or third best tokens made for better output, so we added temperature. And then we started running them in optimized ways where your answer is deterministic only if the batch of tokens are the same (not your input tokens, but other tokens in another batch being processed), and in practice those are never the same. Lastly, we use harnesses that do things like adding IDs and timestamps to the context, which means the same exact text from the user does not lead to the same text hitting the AI.
The final result is that, in practice, you are right (unless you run a model fully locally, where you can seed temperature and turn off all these other features). But strictly calling it non-deterministic makes it sound like the underlying algorithm is itself non-deterministic (and I've seen many people with that misunderstanding) rather than it being a result of how we purposefully changed the algorithm for better results.
A bit like saying path finding is non-deterministic, because having the best pathfinding makes for poor gameplay, so we added some randomness to NPC path finding to make it more realistic. The given implementation is non-deterministic, but the underlying algorithm isn't.
I feel as though you are overlooking simple statistics/confidence intervals. It absolutely possible for two different works to be not have a distinguishable difference in quality.
Agreed, but not if one of them was altered to contain a secret message. That one will have a distinguishable difference in quality. Maybe (almost certainly) negligible, but still there.
Others have already said this, but the watermarking is something like "when the model flips a coin picking between two values, always choose heads". It was already flipping a coin. You're not choosing a less good result, you're just using a deterministic process when it was stochastic before.
This will have some impact on outputs, but unless you have some reason to believe that always picking tails was better than always picking heads (in which case, you should be working at one of these companies in model training!) it won't have any impact on output quality.
It has an effect, and it's negative. It's hoped that the effect is negligible, and it probably is, but the whole point is that it has an effect.