Anthropic's ‘watermark’ text adulteration in Claude is a perversion of writing
daringfireball.net
daringfireball.net
So every check requires sending the text to as many AI providers as offer a watermarking detection API, almost all of which have a very dubious track history with obtaining training data through illicit means.
Any university using AI detection in their submission pipeline, or lawyers, editorialists, proofreaders that check for AI marks will be sending significant amounts of text like unpublished research, books, potentially internal documents and more, most of which is high quality human written, to dozens of AI companies, blindly trusting they won't train on any of that.
I wonder if this would be a good use for homeomorphic encryption. There might be a way to let anthropic check some text without actually giving them access to the source text. Any experts around? We could use your skills!
A good fingerprint should make use of cryptographic signatures. Without knowing the keys, the fingerprint should be indistinguishable from noise (or just random token selection)
They're giving you an oracle regardless, which is almost as good. Take LLM output, make some modification, ask the detector if it's LLM output, repeat until you learn what kind of changes you have to make to defeat it.
Or don't even bother learning what to do, just make arbitrary changes until it says it's not, so when the person they're submitting to does the same check it says the same thing.
Reference needed? I think it remains to be proven whether those detectors can be considered deterministic.
Alternatively they could also just keep saying "yes" if it's close enough to a version that was close enough.. Although that would enable the attack to allow arbitrary text to be "proven" AI, by slowly morphing close-enough generated material to the desired text. But perhaps this is not a problem they are not concerned with.
To satisfy the letter of the law I expect it's enough to just provide the oracle, without any mitigations.
being allowed to train on any data that you can legally obtain ought to be a right for anyone.
After all, i am allowed to learn off anything i can legally read (and perhaps even illegally read). The only thing not allowed (rightly so) is to produce a copy with enough similarities that it can be replacing the original.
Why should we hand over even MORE power to the owner class?
In a fantasy world this could be possible yes.
It's part of why we sign NDAs, and why their duration is measured in years (and that's not even targeting the human retention - just duration after which information ages enough that its disclosure is not likely to negatively impact anyone who cares).
But LLMs are replacing the original, just in different words.
And what does 'legally obtain' mean in this context? Copyrighted content is usually licensed for specific purposes. So if a license is given from training your LLM, then by all means do! But what if the license is 'for personal use'... ?
I have the opposit viewpoint to the extreme. They shouldn't be allowed to even read that data until they are very clear about what they will or not do with it.
Can they publish it? Can they store it? Can they use the information in it on prediction markets? Etc.
Humans reading texts historically come with little negative consequences, but machines reading and processing texts en masse is more dangerous and should be regulated.
Citation needed. This is sounding tautological.
IDK, we do have laws against opening other people's mail. Those have been on the books for hundreds of years. Seems like someone figured out a while ago that certain unauthorized humans reading certain restricted text wouldn't be good.
Correct, if you violate it too often to count, you have to pay around less than ~2.5ct per violation.
So the lesson here is: Create a company to do torrenting professionally, and resell its values for higher prices. Then get sued and pay a dime on the dollar you made.
edit: Actually it's 2.5ct per violation.
(1) https://apnews.com/article/ai-anthropic-copyright-settlement...
Are you a tool?
Because humans gets rights, tools don't.
Arguing that untrained or partially trained models should have have rights is a different argument to arguing that a trained model should get the same rights as a human.
Having proof that content (especially images and video evidence) is unmodified (whether via Photoshop, Paint or a model) is far more valuable then having evidence that an image was manipulated or generated fully by a model (which still leaves other forms of manipulation), I feel the same goes for human authored vs generated text. Free to admit that using models to generate any kind of media whole-cloth is still unappealing to me and I still pay for commissioned artwork or make it with my limited abilities for what that's worth. Do like to (poorly) write my musings too and see UX as something were thoughtful contributors (like the opinionated, sometimes controversial, but certainly talented GNOME Gitlab contributors) can make a major impact.
Code can be beautiful, interesting and serve purpose beyond execution, of course, but for most people, in most cases, it does not in the same way as audiovisual content (not limited to art). Having code just to execute and resolve a problem can have value all in itself, the code being a means to an end whose quality, let us be honest, was barely a concern in most corporations long before LLMs.
Also have rarely (honestly never) before LLMs fully owned all parts of any code base, always relied in part on someone's prior effort in (Flutter/Dart mostly) packages, whereas when writing, drawing, etc. I have far more situations where I make something from scratch and everything there is only there because of my conscious decision. Even simple marketing mockups that, quality wise, any modern model would beat feel different when I was fully in control, where to place what, etc. Objectively worse (at my skill level), probably, but still never the same.
Knowing something was made from scratch by a human has value to me, beyond misinformation prevention. Knowing for a fact that LLMs were used instead of importing a library, using a template, or something similar that leads to expending similar amounts of effort, I don't see that being nearly as valuable. Heck, with all the importing and my experience back then vs now, I am spending more effort actually fully reading any LLM output in my code then I spent back then auditing Flutter/Dart packages. Then again, LLM output fails far more unpredictable then those messy packages that simply got Gradle to take down my system...
Happy to admit, I have been skeptical of watermarking LLM output being feasible for quite some time and having looked into SynthID Text and proposals being researched, I am convinced that it is challenging to impossible beyond the lowest common denominator and less important then proofing human authorship.
It will catch people just copying LLM output into their replies without thought, which is not a negative in my book, especially if it is not discernibly affecting output quality in regular use cases. Anyone who wouldn't copy Wikipedia into their dissertation will, in my opinion, be able to bypass text watermarking as proposed however, I feel we need to be honest there.
Thing is, if that's the case and text watermarking will only ever catch LLM created slop, is that a bigger problem then the misinformation, harm to creators due to authorship questions and accusations, making it harder to use evidence in proceedings, teachers not trusting students even when they did the work themselves, etc.? Signatures for all such cases will be difficult to implement, yes, but I feel are going to be of greater value in the not to distant future and I equally feel are not impossible, not least because idiots will always want to hide their LLM usage, whereas human authorship is something they take pride in and want to proof.
Because a malicious human will gladly copy/paste LLM text and sign it with his "I, a human, definitely wrote this academic paper" key?
I will fully admit that at a point in the future, maybe not too soon, models may be trained to bypass that too, at which point we are back where we started. As a skeptic of the extend that capabilities are emergent in LLMs vs specific to training data, I am somewhat hopeful that unless models are specifically trained for evading such human detection solutions, they'd struggle to do so, but it could still end up as a byproduct of improved, lower latency computer use focused training. Not emergent as the term is used in regard to models because that is still output performance improvements clearly traceable to very specific training data, but incidental as the goal of said training data was not to bypass.
For what it's worth, I find human authorship being verifiable to simply be the more crucial problem over watermarking model output, so if research is to focus on one, I'd rather it the former. Maybe both signing human authored content and watermarking LLM output are both only possible in the near term, I hope not but fear it that might be the case. If so, we as a society will have some major challenges ahead (beyond all the ones we'd have anyways).
Alternatively, we could also just start scanning everyones eyeballs...
Then build our assumptions and how we operate around those trust levels in the digital realm.
Artists, photographers, journalist, etc. are going to want and need this.
But even then, people will be able to point that camera at a manipulated/generated image (either printed or on a screen). Maybe that one could be solved if the photo included some depth information?
I suspect the signal will be significantly under the noise floor, so it's not detectable if you don't know exactly what to look for, but certainly you can submit more information then the textual contents.
After-the-fact checking only needs the vocabulary splitter, which is independent of the LLM. Over a sufficiently large text non-watermarked text would expect to use green and red tokens with the baseline probability, and that difference can easily become statistically significant over sufficiently long texts.
The basic algorithm has obvious knobs to tune, among them the initial ratio of red to green tokens and how hard the sampler tries to pick a green token. These would balance fidelity to the original distribution against watermark detectability (minimum required content length for statistical power).
The Nature paper is "Scalable watermarking for identifying large language model outputs"[1]. This method does not separate out tokens into separate classes, but merely uses a seed for the PRNG that selects which among the most likely tokens generated by the LLM will actually be output. This has the advantage that there's no green and red token sets, so no token is systematically favored or disfavored. If a particular token is overwhelmingly predicted to be the most likely candidate, it will almost certainly be selected, so the watermark doesn't affect that. Even if there are several choices of output token at a point that have similar probability of selection, the watermark doesn't systematically bias in favor of one token or the other.
This is actually a quite elegant method of watermarking that, contrary to people's fears, won't adversely affect the model output. The main concern I have with it is that it appears that you can't actually test the watermark locally, without uploading it to Anthropic. I'm not sure why that's the case, since there's no particular reason the watermarking key has to be private, except if you want to prevent others from generating text with their own LLMs that is watermarked to look like it's generated by Anthropic - but everybody wants their text to not have the watermark.
[0]: https://www.anthropic.com/news/claude-text-watermark#:~:text...
The upside is that this has very low false positive detection rate, but the downsides are many. It only works on longer pieces of text. The system is fragile, and small edits (or rewrites by a local model) can fool the detection. Only the owner of the model is able to re-run inference at this level, so data must be sent to them for evaluation. And sometimes the token output is basically 100% deterministic because the input asks for the straight answer to a fact, or to recite a quote verbatim. That leaves no room for watermarking at all, unless the model is able to lie.
That's a fine place to put the watermark to track those users who accept the code blindly and don't delete/edit the comments.
In the case of Anthropic it would also be entirely unsurprising if they've been lobbying the government to force everyone to do something in their (Anthropic's) own best interest.
It’s also in their interest to demonstrate that they can be trusted and to show that they at least pay lip service to limit the obvious downsides of the tools they are selling. The use cases they sell to mainstream audiences are not affected by detection tools. The point of having a LLM do the work for you is that the work is done, and reliably. It does not matter if it is done by a LLM, and most of the time it is obvious anyway.
So, the use cases that are being sold to mainstream audiences actually will be affected by detection tools, especially if the output is intended to be monetized in some way. In the near future the EU will likely come down with heavy intervention to prevent AI from impacting employment rates across Europe. The number of legitimate use cases for costly frontier models drops significantly once eliminating professional jobs is off the table. This is all conjecture at this point though.
> The EU mandate for AI watermarking is likely the first step in the direction of prohibiting AI for specific use cases.
I am not sure how practical that would be. The cat’s already out of the bag and they won’t prevent companies in the whole world from releasing open weight models. Playing catch up by distilling flagship models is also relatively cheap; we’d see smaller companies setting up shop in friendly regimes. And I don’t see any appetite to go full child porn and criminalise the possession of a LLM. So we’d end up with a similar situation as with illegal downloads, i.e., everyone will do it and nobody will care.
> The pretext for outlawing (or at the very least controlling) the use of AI when the time comes will be something along the lines of data integrity or just general compliance legalese.
They could forbid using LLM for hacking, but hacking is already illegal. They could make it a factor when determining punishment, but I don’t think that would work terribly well. Most of the dangerous stuff we can do with LLMs is already illegal, or should become so. Things like propaganda, identity theft, harassment, scams. We need enforcement with teeth on these, not pointless feel-good legislation. Again, there are parallels with cryptocurrencies and torrenting software. These things have illegal uses, but it’s also really difficult to make them illegal, at least in semi-functioning democracies.
> So, the use cases that are being sold to mainstream audiences actually will be affected by detection tools, especially if the output is intended to be monetized in some way.
I don’t know that mainstream audiences are really against LLMs. They are mostly against AI in a nebulous sense, but even non-technical people use ChatGPT or equivalent. I think that the critical mass is already there and the tools are convenient enough that they couldn’t outlaw them without a massive uproar.
Detection tools don’t seem all that relevant to mainstream audiences’ use of LLMs. AI companies will sell this as a safeguard against misuse, and everyone will be happy about it. The politicians will say they accomplished something, the AI companies will slowly turn public opinion, and the public will have shiny toys.
> The number of legitimate use cases for costly frontier models drops significantly once eliminating professional jobs is off the table.
I don’t know. They can open possibilities that we don’t necessarily consider.
One example I have is a friend who is getting his house refurbished. He’s not an engineer or a material scientist. He does not have enough free time to read thoroughly on the many subjects involved. With a decent LLM, he could untangle the technical documents sent by the architect and the contractors to really understand what was going on and be involved, rather than passively follow the architect’s advice. For starters, the LLM was very useful in finding issues in the quotes he received when he was looking for an architect. Those were long, technical documents, with no really standardised structure and full of jargon. I don’t think that person is going to want to stop using LLMs now. Many people are having this sort of moments right now.
> In the near future the EU will likely come down with heavy intervention to prevent AI from impacting employment rates across Europe.
Maybe. But i don’t believe the EU is well equipped for that. Labour laws are largely local and different in each member state. The EU regulations are basically the common denominator, ore or less, and it is easy to see why: for regulations to get adopted, they need a strong enough majority in the Commission, in the Parliament, and in the Council. It is very difficult to get anything controversial that affect the sovereignty of member states passed.
The angle of the current AI regulations is that they set the rules for the single market, which is where the EU is the most legitimate. It is difficult to see market angle for the effect of AI in labour, and I think enough member states would be keen to kill the project.
Also, there are many influences at play, but the EU is fundamentally an economically liberal institution. It very rarely goes in the direction that reduces economic activity. Look at how clumsy it is at fighting against cheap Chinese imports. I don’t think the institutions themselves would really want to make AI illegal. Companies have too much to lose.
I think your comment about torrenting touches on an interestingly relevant case study, specifically media piracy. The seed of that technology was sown when the internet was still just clusters of machines passing files around and then exploded with the PC, Napster, and TPB. Napster tried to be a legitimate commercial enterprise with a business model of undermining the ability of copyright holders to rent-seek on the consumption of the material they 'owned'. At the core it was a novel technology (P2P) that revealed an economic arrangement to be out of date (if a distributor no longer has to manufacture a copy of the media for each individual consumer, their business model boils down to rent-seeking). Western legal systems were quick to rule on the matter (in favor of copyright holders), and I have no doubt that they were looking ahead to a future wherein media creation was totally disincentivized by said novel technology.
Now a novel technology (the cloud inference-backed LLM) is challenging another economic arrangement. This time around, the arrangement being challenged is the higher education->professional job pipeline. All advertising and media messaging aside, it really does seem like the frontier labs are only economically viable if they get massive enterprise deals across a broad spectrum of industry. There is fundamentally one chunk of capital organizations are going to spend either supporting their talent pipeline or padding it (to put it gently) with enterprise LLM deals. If the latter path is taken too far, consumer spending plummets (due to lack of middle-class incomes), assets backed by consumer debt/spending fail longterm, and we will have to deal with a deluge of socio-political issues stemming from the absence of real social mobility (we are in the early stages of this now, incidentally).
All this to say, I think parallel situations from the past can guide our thinking re: AI regulation and the forces shaping it. Reasoning based on regular political/economic incentive structures (e.g. economic liberalism not wanting to reduce economic activity) will fly out the window at lightspeed once fear becomes a factor. I personally think its great that LLMs empower individuals such as your friend to increase the control they have over real issues in their life; that is what technology should be doing for us. Cloud inference-backed LLMs are doing the opposite: drastically reducing the power that the everyman has over his socio-economic future by throwing high-paying career paths for a whirl and incentivizing powerful organizations to destabilize the labor market. Only time will tell how this plays out.
But unless you know the prompt, you don't fully know which choice the model faces. Surely a lot of coding space is wasted compensating for that uncertainty. Exotic prompts ("Use no more than five E's in four consecutive words anywhere in the text") seem like they will confuse the hell out of attempts to extract the bits from the output text alone.
When you only need to encode one bit, the signal to noise ratio can be very low. If I try to write my own human words under the policy of "try a little bit to avoid the letter 'e' in every fifth word," then a sufficiently long text would still be 'watermarked' even if I only succeed in this dictum (e.g.) 10% more often than the baseline.
A few caveats: you need to know the key to the function (used when generating the text) and you need to know the bias it would introduce
The point is that you do not need to know the full prefix, just a modest sample set of contiguous tokens
Well no, small edits wouldn’t fool the detection as long as the seeding only uses a small run of previous tokens.
And yeah full rewrites breaking it is by design. The watermark is just meant to tell you whether the text was generated by a watermarked model, not whether the ideas came from AI or something like that.
>use it to store arbitrary information
No additional data is embedded. The range of available data is constrained by the text being generated (i.e. the sets of "next words" per text).
I attempted to clarify that the impositions themselves are not deterministic, by indicating that the entire process is still probabilistic.
Maybe Anthropic's explanation is simple enough [0]:
>When watermarking is used, choices are still made at random, but the source of the randomness is different. Instead of using an arbitrary random number generator to pick the next word, watermarking uses the key and a few words that come before to settle what word the model should pick. That is, the words that Claude picks are still random, but now, one can check the sequence of words and see if it’s consistent with the choices Claude would make if it was using the key. If it is, one can assign a probability that the text was generated by Claude.
It’s quite fun and engaging.
If you care to avoid detection, yes, it is useful. If you care about the best possible sequence of words, then the damage is already done once watermarked.
its also kind of laughable that somehow people are trying to prevent the outputs not to be altered. Asif you cannot manually paraphrase anything you can read. So the only solution would be, to make it utterly unreadable (which is not possible, it obviously defeats the purpose of the thing).
Not to mention local models ofcourse :-)
I mean, all of these text content watermarking schemes require the company to assess if the text was AI generated or not. They aren’t going to tell us where the toss-up tokens are or what is in the red vs green pools of words.
I read stacks of term papers all year and it is a reality that, apart from such schemes, we are in an extinction event for civilization.
I would think that these things would eventually converge and we’d get one watermarking algorithm as an industry standard. That way, all major provider would follow it and we’d get independent software for checking. This would partly limit the efficacy of the watermarks, but on the other hand if it’s done correctly, removing the mark could still be enough of a pain that casual users would not bother. That would obviously depend on a lot of factors. It would at least add significant friction in the production of daily slop.
Of course it wouldn’t do much for thing like foreign propaganda but that’s a whole other discussion we need to be having.
> Any university using AI detection in their submission pipeline, or lawyers, editorialists, proofreaders that check for AI marks will be sending significant amounts of text like unpublished research, books, potentially internal documents and more, most of which is high quality human written, to dozens of AI companies, blindly trusting they won't train on any of that.
Isn’t it already what they are doing right now with some of the plagiarism detection tools? Not every university is going to have a representative corpus, and yet they are all using the software. So I guess the provider is doing the work of feeding all that data to their algorithm.
Couldn’t it be checked in the TEE using confidential computing to keep Anthropic’s algorithm secret?
Uh, yeah, why do you think it's setup this way? The frontier companies desperately desire more high quality human text and this is how they are planning to get it for free.
Local models could probably catch some patterns.
LLMs cause a wide number of problems, the good news for our investors is that they're all solvable with LLMs.
Then bad news: LLMs already use randomness in a fundamental way. Each time they go to generate a token, they first generate a probability distribution of possible tokens. Then they pick one randomly according to this distribution. The technique described can be thought of as making the random number generator pseudo random. The output it generates is one of the possible outputs it would have generated before, just now it's deterministic and will generate the same thing every time.
(1) The behavior that is approximately what you describe is not "fundamental" (though it may not be something you can disable on some hosted providers), it is an option that is not fundamental (and with runtimes where you have full control can be either disabled or tuned in a large number of manners), and
(2) The actual behavior that is approximately what you describe already usually involves use of PRNG (with a user or harness supplied seed), not a true RNG; the change to do watermarking isn't going from RNG to PRNG, it involves adding an additional set of constraints on token generation on top of the existing ones, which inherently compromises quality.
Related: if you don't have a limit on sampling (top-K or top-P), eventually you'll hit one of the really unlikely tokens by chance and then the model will switch to Japanese because the most likely completion after a random Japanese character in the middle of an English sentence is more Japanese writing, not a reversal back to English.
LLMs are likely to get stuck even with sampling if asked to generate tokens on their own long enough, though sampling does tend to stretch out the time before that happens (as do other techniques that don't involve sampling, like applying repetition penalties directly to token logits). But LLMs generally aren't left to infinitely extend their own output, and the length response typically needed in the use case is much shorter than the would result in collapse given the kinds of inputs expected in that use case, the existence of the theoretical eventuality may not really matter.
I don’t see how this follows? Tokens are chosen randomly. If you choose tokens with a different RNG in the same distribution, you’re still getting equally good or bad tokens.
Writing has rhythm, or at least it's supposed to, and synonym swapping compromises it.
Never mind metaphors and similes, which are even more tightly constrained.
LLM writing is still a long way from good. Sometimes you get lucky with the odd line, but there's a difference in quality between influencer slop, genre fiction, and literary fiction and/or best-in-class journalism.
LLMs are still somewhere between the first two, and nowhere close to approaching the third.
We already know that a non-zero temperature improves quality though with current models (particularly with creative writing). The assumption that always picking the 'best' token results in the 'best' output is not the current reality.
And if you are already intentionally putting in randomness, I can imagine that it would be possible to seed the randomness in a way that is detectable but results in the same quality.
This is obviously not true for queries where temp = 0, but at temp = 0 then it becomes easier to identify anyway. I assume this technique implies some level of temperature.
2) Claude’s PRNG having a P is immaterial
In case of LLMs, you can look at it from high and low level.
At low level - if you could do with less randomness, you can always lower temperature. You usually keep it (or for SOTA providers' chat UI, they keep it) at a level where it's about right level - high enough to allow for more creative leaps and interpretations, low enough that it doesn't go off into crazy land after the third paragraph.
At high level - creativity is driven by randomness. If you had an author (fiction or nonfiction) you like for their both broad and deep range of insightful thoughts, would you be happy if they suddenly developed an acute porn obsession and uncontrollably added lewd subtext to every other sentence? Still creative, still deep, but now with that one strong attractor that biases their every thought in a single direction? Would you trust/enjoy their output as much as you did before?
That, slightly exaggerating to make it more obvious, is what "loss of quality" means here.
I guess they can at scale filter the watermarked documents (by necessarily allowing eachother to at scale checked for watermarks, but banning the labs not part of the watermarking-cabal). Makes me wonder how useful the human quality filter is on AI output - if a human judges a given output as genuinely good and posts it somewhere for the scrapers to find and take into the training sets, will these types of outputs also be filtered out?
2. (raw, pre-watermarked) Output token probability situations where 1 output token has the majority of the probability mass associated with it, but it is not in the watermarked set, will force the model with much higher probability to walk a non-optimal latent space. E.g., if the next OBVIOUS token for a given sentence would be a point, but the model is in this way not allowed to output it, it might put a comma and start off on a whole different tangent just to make the initial non-optimal comma grammatically make sense.
You're missing the fundamentals here!
Random watermarking functions colour the tokens based on (small) contexts and a secret key. Given watermarking functions are randomly chosen every single time they are used (so essentially not deterministically seeded by the context and secret key), then indeed the completion distributions are unchanged. However, they are deterministically chosen, a given string of text will always have the same corresponding watermarking functions. Tokens scored high by the function see an increased probability of being the chosen completion, those scored low see a reduced probability. I dont see it is different from merely talking about it as green/red and the points hold?
The only way you'd notice this is they weren't independent, and the most plausible way that happens is if you're re-completing pre-fills (resampling the same hash function against the same log-probs).
Anyway, there are lots of cases where text carries very little entropy. E.g. boilerplate code, exchanges of pleasantries, well-worn platitudes and jokes, etc. These are sequences of tokens that will be seen across many, many separate outputs. Watermarking here (on the token following the common sequence) would thus be easily detectable and noticed as a claude style. The longer the hashing context though, the lower the amount of pathological cases with low entropy. It would be interesting to understand the exact parametrization better!
The real question is: "What can manipulation of pseudo-random number generation do?"
We know that in the cryptographic world, attacking "randomness" is a key offensive capability. It will be here as well -- if Anthropic can watermark text as generated it's LLM, will it be able to watermark outputs as generated by "Spooky23/FooCorp"? Can I pay Anthropic to steer inquiries in a way that benefits my company or governemnt?
Pseudo-random to the end user appears random. Most treat it like a random chance. It is not.
On the other hand, LLMs are forced into picking some likely-ish word, and then have to build the rest of their response to retcon that choice into making sense.
Even good human writers would probably struggle with this constraint. It would be like someone interrupting your writing to tell you the next word MUST be such-and-such, and then you have to try and make it work as best you can first try, without going back to edit. The result would probably be a little clunky. (Maybe it’s impressive LLMs write as well as they do.)
But what about the general idea that they can watermark results to tell where they came from. The next step is tracking down which user got a result. I hate both of these things. Must everything we do be tracked? Next altering wikipedia results so they can tell who looked at the page or something?
I'd like "the best answer" from an llm and don't want to be tracked, but this isn't for me, it is for them. I understand llm results are already using a varying statistical input so they aren't always the same. But I really hate watermarking and likely tracking too.
Here’s an example: I had asked Claude for some music recommendations in a certain style. Part of its output was:
—
*Long journey tracks*
Clinic — “The Return of Evil Bill”
Guided by Voices — not really, wrong band
Silver Apples — “Oscillations”. Proto-everything, deeply repetitive, hypnotic.
—
So at some point there, the next token produced was “Guided” or “Guide” or whatever, and then because it can’t go back, it had to correct itself after the fact.
Reasoning/CoT have helped a lot, but I feel like small versions of this still happen all the time.
Human writing is like 90% editing.
Reasoning tokens with tool calling tell the model to loop on a one phase of a question and call a tool to indications completion when done.
Related, but not the same thing.
Presumably you could use the same reasoning trace, run multiple generations, and get different outputs (if the temperature is >0).
But now I’m interested in playing more with Cowork or Claude Code/Codex for prose writing to see if the set of tools there affects outputs at all. I guess you might need a more custom “writing” harness.
But also, no one really knows as they're closed.
I've heard of attempts to use diffusion models to generate text or code as well, but my impression is that it just didn't yield the level of results necessary to dethrone a frontier transformer model.
Mathematically, a long chain of conditional probabilities is equivalent to a single probability over the whole range. But computationally, for that to work out, the computation for the first probability needs to somehow consider all the downstream probabilities depending on it, which obviously isn't how autoregressive language models work. They can pack in as much downstream computation as their neural architecture allows for, which is quite a lot.
Suppose in some context you have three equally plausible conpletions after "Be": "tween a rock and a hard place", "twixed he stood there" and "lieve he can fly". To model this probability distribution of the whole sentence, the next token "tw" needs to appear at 2/3 probability and "lie" at 1/3. After "tw" would be a 1/2 chance of "ix" and a 1/2 chance of "een"; after "lie" would be a 100% chance of "ve " and in any case the rest of the sentence after that would be 100%.
The model needs to somehow "think ahead" to know those are the possible completions. For example if "lieve he can swim like a dolphin" was another equally plausible completion, that first token would need to be 50/50 instead of 67/33. So the computation of the first token somehow needs to encode the fact that the guy thinks he can fly but not swim, even though it doesn't become relevant in the output until several tokens later.
In practice this probably happens to some degree but definitely doesn't happen perfectly. To perfectly model the first token's probability distribution, it would have to include knowledge of the entire distribution of all possible outputs, which is just not happening. So it approximates. Surprisingly, the approximation is good enough to produce language.
You can see this breaking down in the seahorse emoji incident from last year. When you ask the model if there's a seahorse emoji, it first completes "Yes," as if a few tokens later it's about to produce a seahorse emoji. But when it actually gets to the token that would produce a seahorse emoji, it can't because there isn't one. But it's already outputted "Yes, the seahorse emoji is" and can't just go back and change that to "No, there's no seahorse emoji." Some models would try a few times and then say there isn't one or a system error seems to be making them unable to produce one, other models (including then-current ChatGPT) would loop forever with ensuing hilarity.
You're mixing up two claims here, and only one of these is kind of true. Yes LLMs do internally plan ahead in a way that is emergent rather than strictly part of their architecture, so that part of your claim is true. The way you word it by saying they are "coalescing the probabilities of a range of tokens at a time" is poetic sounding jibberish though. What's actually happening is one distribution output for the next token computed from a hidden state that implicitly encodes where the text headed.
Your claim that if an LLM does happen to pick a token "th" instead of "tw", then the LLM isn't stuck with that decision is entirely false for autoregressive LLMs which is what all of the frontier models are. Whatever an LLM picks as its output token is final, it has no ability to undo that token selection and it must continue on the basis of that choice. It can't go back on that decision and revise the output.
If you're interested in this, Anthropic has a summary of a very technical paper on this topic that mostly deals with this issue with respect to poetry:
https://www.anthropic.com/research/natural-language-autoenco...
So we train a second copy of Claude to work backwards—reconstruct the original activation from the text explanation. We consider an explanation to be good if it leads to an accurate reconstruction. We then train Claude to produce better explanations according to this definition using standard AI training techniques.
Well, I suppose it's nearly the opposite of that experience, upon further review. But for some reason, that's where my head jumped.
You be a human who's brain shifted into LLM mode (chainneling Markov?).
Or perhaps you're an LLM impersonating humanity.
I often wonder how much LLMs are just mirroring our own brain's patterns.
https://chainofbranches.com/conversations/2/branches/20/
I’m not convinced it’s possible. A good nights sleep and a notepad in a quiet room still feels like the state of the art toolchain for writers.
They are also usually worse (which is often better!) because they are usually lazy and don’t want to spend effort they do not have too, to accomplish their goals.
Their goals are often complex and nuanced.
None of this is true of LLMs.
I feel like the existence of good writing is also not impossible but not very likely, and so of course LLM can only write mediocrity, even when taught only on great writing.
But that would be a fun writing exercise, I think. Thoroughly in the oulipo wheelhouse.
Maybe generate a Markov chain table over all of Project Gutenberg and then say every 10th word is whatever the Markov Chain thinks it should be at that point?
Or every Nth word has a P% possibility to be constrained by the chain? Optionally with the possibility building for each skipped word to guarantee it happens at some point. Bonus with this approach is that the human can't game the words leading up to the constraint because you don't know when it will happen.
That's what LLMs in reasoning mode do, too, to the text they present to you.
Even so, I don't think it will stop here. Once this is in place, the next step is to put more and more identification into the AI generated content; might as well pack it in, it's not that bad, and if it is they won't admit it. There's no way for anyone to check. And your argument will still be technically correct but missing the point.
I think there is a possible weakness in the context of the watermarker but that is not your claim iiuc.
Could it have been equal or better with slight variations in wording?
The slipper slop argument is too lazy to address directly. Argue A is bad because A, not because A might become B and you’ve got good arguments against B.
In fact we know it's not that good because we can often tell Claude's writing apart from human writing.
> To illustrate, in the special case that GPT had a bunch of possible tokens that it judged equally probable, you could simply choose whichever token maximized g [a cryptographic function]. The choice would look uniformly random to someone who didn’t know the key, but someone who did know the key could later sum g over all n-grams and see that it was anomalously large. The general case, where the token probabilities can all be different, is a little more technical, but the basic idea is similar.
Seems like - given enough text to encode information into - it would be possible for OA to uniquely identify the user account (and maybe even the specific request) that generated some content, even if the chat text itself isn’t stored.
Interesting argument in favor of local AI as a mechanism for privacy-preserving generated content. Though, I wonder if there’s a way to “bake” a hardware fingerprint into local models as well…
Watermarking changes the probability calculations for reasons other than quality. It can't not compromise quality. It literally leads the LLM to occasionally chose different tokens just for watermarking purposes.
Also, as long as the same sampling strategy is used during training as the one used during inference, then the LLM will actually do much better with the biased sampling strategy than it would with a fair one - because that is what it was trained to optimize.
So what? By definition with this system the LLM will chose tokens it otherwise would not, purely for watermarking reasons. Yes this token may have had a decent likelihood of being chosen anyway, but it wouldn't have been chosen and now it was for reasons nothing to do with output quality.
I'm not sure what your last paragraph is trying to say. The blue/green list system changes what output the LLM would otherwise produce. You can't train it to produce watermarked output with this system. If you tried to, there would be no delta between trained output and watermarked output for you to be able to detect.
My second point is that the training of a model by definition maximizes the fitness between the final output function and the training metrics. So, if the model is trained with the watermark applied, the training process will minimize the function `model_error(input) = |watermarked_sampling(model_output(input)) - expected_output(input)|`, by definition. This means that a model trained in this way will perform better when sampled using the watermaked_sampling method than if using, say, top_k sampling.
>My second point is that the training of a model by definition maximizes the fitness between the final output function and the training metrics.
Right, but the fitness in question is watermarked text fitness, not fitness for any user interests aligned metric. You're basically saying that if we train LLMs on watermarked text they'll be really good at producing text that looks watermarked, and then we'll stick an actual watermark on top of that. Screw whatever the user wanted it to be good at.
Yes, that's the goal that was used, but they are quite simplistic and crude methods, not some specifically designed function, with carefully fine tuned parameters or something. So, if a basic function like top_k can improve model utility, it's not impossible to imagine that watermarking could also happen to do so, or at least not have a significant negative effect. So whether the effect is deleterious or not is an empirical question, not something we can assume ahead of time.
> You're basically saying that if we train LLMs on watermarked text they'll be really good at producing text that looks watermarked
No, you're misunderstanding how the training works. If we train the model's output so that it minimizes the error function after the watermark is applied on it, the model will learn how to produce the best output it can given the watermark. It will produce better text that happens to be watermarked, not "more watermarked text". Same as if you train the model on minimizing `top_k_error(input) = |top_k_sampling(model_output(input)) - desired_output(input)|`, the model will learn to produce better output under top_k sampling, not learn to produce output that's "looks more top_k".
For practical purposes it isn't.
We already use an RNG at inference precisely because it leads to higher-quality output. Changing what function is generating our random numbers changes the sequence, not the randomness from the point of view of a user.
Fundamentally the article is railing against --temp > 0.0. He doesn't know what he's talking about.
You can't as a user tell by how much the quality of the output was degraded. True.
>We already use an RNG at inference precisely because it leads to higher-quality output. Changing what function is generating our random numbers changes the sequence, not the randomness from the point of view of a user.
I'm not saying it wasn't random and now it is. I know how these things work. I said that the quality of the system is in the quality of the probabilities. That quality is being degraded.
How is it being degraded exactly? The probability that it picks each option will still be the same, just deterministic based on a seed generated from the text.
LLMs already use PRNGs. This is just changing the source of the seed. And a different seed does not change the "quality" of the random numbers. Even if you are worried that it somehow might, they can just use a cryptographic PRNG, then it is literally guaranteed that the source of the seed will not affect the output in any noticable way.
Let's say the LLM is in the middle of text generation and "decides" that the next token is "dog" with p=0.55, or "cat" with p=0.45. With a temperature of 0, the model always picks dog, because it's the most likely next token. With a temperature of 1 the model picks dog 55% of the time and pick cat 45% of the time.
With this watermarking scheme, the model might alter these probabilities s.t. p_dog for this particular generated token goes up or down. Let's say it does down, s.t. p_dog is now 0.45 and p_cat=0.55. Now, with T=1 the model picks cat 55% of the time and dog 45% of the time. Regardless of whether the "watermarking function" raises or lowers p_dog, the probability distribution for this token has changed, and whatever math this trillion dollar company and its brainiacs came up with to decide that p_dog ought to be 0.55 has been "adulterated". As others have mentioned there is no way around this.
---
Regarding the watermarking scheme, it works because it doesn't just alter p_dog for this single output token. It alters probabilities for many of the generated tokens (it could do this to all of the output tokens; it's an implementation detail). E.g. at token N, it favors "cat", at token N+1 it favors "house", etc. This way, if you have the secret key that lets you generate the watermarking function for any output token, you can analyze a run of tokens and check whether it's likely they were generated according to your watermarking scheme. The longer the run of tokens, the more certain this check becomes (it becomes extremely certain quite fast).
Then write your own damn text if you care about the exact wording so much
Watermarking per model is just the start. The method is cheap enough to distinguish individual users.
Then use pen and paper. It is the same, you say, right?
If they are adding so little value as to be as transparent as a pen and paper then why use one at all? Transcription doesn't need an LLM so that's not what you're taking about I assume.
That's their whole damn value prop: outsourcing thinking and producing without understanding.
I don't need to read emails in detail to respond any more.
The problem is, that LLMs worked very well for me to improve my writing. Especially as I'm not a native speaker, it was a great way to improve the legibility of my work.
I want a tool that helps me improve my writing. A tool I can learn from. Not a tool that switches out "bananas" to "airplanes."
I'm was using Claude Opus and now Fabel extensively for editing my texts and I find the recent updates abysmal. Not sure if it's due to the Text Watermark.
Before Claude was great in sharpening the meaning in my writing, it's now close to unusable.
The only difference here is that Anthropic is actively trying to make the watermark undetectable.
I was frustrated at first not knowing what they are doing.
After I read this post, (seeing they introduced it in August) I think it has to do with the watermarking.
Try it on a paragraph … the connections between sentences feel clunky now.
I will play with it more and see if that’s really the case (the watermarking making the text edits worse).
The power of confirmation bias…
> Try it on a paragraph … the connections between sentences feel clunky now.
We might have a definitive explanation at some point, but there are about a dozen possible reasons for something like this. For starters, is this something really significant and not something you notice because you are looking for it (again, confirmation bias)? A bit like some people still lose their minds when they see a dash, even though statistical analysis showed that they are not a significant marker of AI-generated text?
This is not about em dashes …
This comparison is frankly absurd.
When writing in English though, I use it more like a dictionary. If you want to write past a certain level, an LLM works better as a metaphor and idiom search engine.
I also like to ask it to generate 20 ways to say the same thing. It’s a great way to simplify or smoothen sentences without losing your voice.
I know that most people don't care, but my online presence is a search query for interesting people, so I care about what I put into it.
I don't expect the LLM to read my mind. The unit of work is too small for intent to matter, and I'll just steer the next recommendations in a direction as needed.
Most of the suggestions are crap, but they can contain the seeds of a good sentence.
Nitpicking here in a way that I would usually avoid, but it is relevant to the conversation being had and it seems like you might appreciate the information... "Readability" would be the more correct word to use here instead of "legibility".
Legibility is close enough for me to know what you mean based on the context, but it really applies to the visual presentation and how easy something is to read at a symbolic level (whether someone's handwriting or font choice is good or bad impacts legibility, whether someone uses good grammar or not impacts readability).
You make it sound as if that were a bad thing.
This cannot change the "voice" of the LLM. It was already letting a random number generator choose which adjectives to use. Now that random number generator encodes a tiny signal.
But fundamentally the way it writes has not changed.
It's not making it choose different words. It's a minor change to how it chooses between multiple nearly identical words, where in the current case it literally flips a coin.
If you want precision and clarity of your writing, then you need to hand write it. Just like when you are optimising, its common to drop to a lower level language because the compiler doesn't express what you want. Sure its hard, but you know, thats kinda the point.
Even if you don't want that, the LLM is an average of the style it was trained to give out. Which is a homogenisation of the language to create a vague padding medium between a few generalised facts. (because a. it makes it less jarring when stuff is wrong, because its smeared over a higher amount of text and b. it looks more 'professional' because American business English is all guff and no meat)
Also yes, two isolated phrases may have subtly different meaning, frankly, the nuance is missed on most people. If you look at the interactions on here, at least 25% of the arguments are caused by people angrily reacting to the things _they_ thought the other person was saying, rather than what the actual person was saying.
So no its not a perversion, the LLM is, if you're gonna be picky about things.
exactly. in the same way that printed books affected word choice, so did the radio.
These were mediums _through_ which communication happened. Language models, large or small, are not any such medium.
Ah my friend, you are about to fall down a rabbit hole into standardisation of spelling, and the sometimes deadly debates about how to translate latin into the vernacular.
English, as she is written, is a great example.
for the spoken word, BBC/received pronunciation is another. I speak the way I do _because_ of BBC radio. The reason I have the accent I do is because I changed it to fit what the BBC put out, rather than what my local (impenetrable) dialect was.
You have to remember that your language is shaped by those around you when you are young. So if you are in an insular community, it will be reflected in your language. If I was a journalist, or hell, just me, I wouldn't be letting an LLM speak for me. So the bastardisation of my voice is down to me, not the machine.
but again, your argument is against LLMs and globalisation of culture, not finger printing.
> again, your argument is against LLMs and globalisation of culture, not finger printing.
ABSOLUTELY NOT. Do not ever put words in my mouth. My argument is that radio is a medium through which humans communicate. An LLM is not.
In the same way that handwriting conveys more information about the writer than type, typing ones own thoughts conveys more information about the writer than prompting an LLM.
The analogy here is hiring a speech writer to do your speeches, or dictating to a skilled typist.
I understand the vociferousness in push back
Technology mediates (human) agency.
[2] > My argument is that radio is a medium through which humans communicate. An LLM is not.
KaiserPro's argument is that radio (a one-way medium) mediates how humans communicate (phonetically), thus influencing how people speak.
Now that does not explain how the printing press and radio have influenced word choice (both of which I would like to see examples of!)
[3] > I can't remember any radio determining words or adjusting grammar of the person speaking through it.
Now media themselves did not really have something that looked like the kind of pseudo-agency that LLMs seemingly have. There may be some kind of qualitative leap.
People developing a para-social (pseudo-social?) relationship with a corporate robot have far bigger problems than the word-chooser in their robot "friend".
I also do it with a LLM every now and then and, while it’s feedback is more useful than a toy’s, it does not need to be intelligent either.
Right. If the exact words matter, using a non-deterministic LLM is a terrible idea in the first place. I hope these people never try putting the same prompt in different sessions.
Also, now I am curious. How would these people interact with other humans? Is there anyone on earth who would provide the exact same reply, down to every single word, if we asked them the same question more than once?
> Marks will apply to output from supported Claude models across Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, and wherever Claude is offered.
I've never asked an LLM to generate writing I wish to post as my own. I don't understand why we think it is a good idea to fudge all output just so people can't cheat on their homework or generate slop. It won't have any impact on those things because there will always be models that don't do this. All it will do is increase the rate of false negatives.
It is deeply misguided regulation and Anthropic should have just said no on grounds of common sense.
(For the record, I read DF quite often as a Mac-minded techy weirdo, I just happen to disagree on this particular issue)
http://blog.tyrannyofthemouse.com/2026/04/open-ai-strikes-ba...
I’m willing to venture that Gruber is on the list of folks that get to hold the opinion choice of words matters.
Cognitive surrender.
The "problem" is that seeing the watermark doesn't mean that the person claiming to be the author didn't make extensive changes to the output of the LLM, or that the LLM wasn't simply the final editor of something that the author had put a lot of work into.
> Cognitive surrender.
I don't know what this means. It's just drama. Don't let the LLM write for you and this is not a worry. I'm not worried about the poetry of LLM output being subtly adulterated.
How does that follow? AI-generated text is already not a perfect emulation of human writing. There's lots of room to affect it laterally without changing the level of quality.
As I understand it, LLMs with temperature >0 can select from many possible outputs. All they're doing is limiting the possible outputs to ones that contain this pattern. I don't see any reason why the quality of that subset should be lower than average. The very best outputs will likely be eliminated, but so will the very worst.
If you get your random numbers from a cryptographic PRNG, then to notice the difference between that and 'real' random numbers even in theory, means you need to break the cryptography. In practice, your gut feeling about how good some text is won't break modern cryptography.
I'm skeptical that anybody generating LLM text is really all that concerned about optimal word choice. Or even particularly good prose. But let's pretend that person exists.
If that person tried, say, an open model and that same model with watermarking applied, I'd be eager to hear their thoughts on the prose quality. Especially if they built an experiment harness and rated a few hundred blinded examples and found a measurable difference.
But getting this upset in advance of any demonstrated problem? It really seems to me like the point isn't the point
The easy thing to do here would be to have 1000 questions, randomly assigning one half to an LLM with a watermark, and the other half without. Then show people pairs and say, "Which one seems watermarked?" (Or, "Which text seems more natural" or "Which is a better answer" or something like that.) If they come out equal, the watermark really is indiscernible, at least to most people.
"Which diamonds are shinier, the blood diamond sourced ones or the ethically sourced ones?" ... that's not the same question as "which diamonds are blood diamonds" (to employ an extreme analogy)
Concluding that no one could detect which ones were blood diamonds because they were "equally shiny" is not really correct now, is it?
And the Daring Fireball article does complain that watermarking will reduce quality. If that's what you're trying to check, "which is better?" is the right question.
Also, depending on what I am asking it, I often don't want to use the thumb down or up, as this may mean my conversation is going to have some kind of human review and depending on what I am asking for, I may not want to bring attention to my stuff.
From what I read, two big problems for long-running columnists are getting tire of the work and running out of things to say. I have no knowledge of Gruber, but I can certainly see why people who are expected regularly to have something to say would turn to "AI". It doesn't get tired and is always ready to spew infinite words.
Yes, but that's neither surprising nor a reason to dismiss the anger. People get angry about DRM schemes in video games, even if the slowdown these cause is practically imperceptible. They're angry -- and Gruber acknowledges that factor too -- because a stranger manipulates what they regard as their own domain, without consent by or benefit to the owner.
It might be another instance of consequentialism vs. honor ethics. Many consequentialists don't seem to understand that something that doesn't have demonstrable consequences can still have moral implications.
Any potential "slowdown" doesn't even come close to making the list of top reasons people get upset about DRM.
> because a stranger manipulates what they regard as their own domain, without consent by or benefit to the owner.
It's LLM output! It's not your domain, it's the LLM owner's!
Yes but thats a thing that degrades something in a catastrophic way, as in I can use the thing one day, and not the next.
A different randomisation system on something that is a text generator which is designed to be unperceptable sounds like the people who are annoyed at FLAC vs MP3[1]
Done right you won't know the difference, done badly and you will.
[1] ex audio engineer, try me.
Good luck explaining that one
What’s the story there? I didn’t know that was a thing and I’m curious to learn more.
Flac takes a raw .wav and effectively zips it up to shave off a certain amount of space. (there are nuances, I think the compression scheme is designed for streaming.)
mp3 is perceptual, so throws away the stuff that humans can't hear. This yields a much smaller file.
However its all a sliding scale like PNG vs jpeg.
a .jpg with a quality setting of 85 will be almost identical to a .png in visual quality. However if you then edit that jpeg, the image degrades and you start to see artifacts. (hence why memes look like shite as they get older)
Its the same with mp3s if you compress the hell out of them, say 64kbit or lower adaptive, then you'll start to hear the tell tail "schlop" noise of mp3-like compression. You might notice it most with cymbals in drum kits. cymbals are wideband noise. as in there are loads of constituent frequencies so if you remove some of the "hidden" frequencies you tend to notice, so they sound more metallic, ironically.
But, all of this is solvable, 256+kbit is more than enough, bonus points for higher sampling frequencies. (however you need a decoder that can actually do that sample rate...)
To the extent that it's anybody's, it's either Anthropic's (they run the service) or everybody's (in that we created the content it's remixing). Legally LLM prose isn't copyrightable for good reason.
that's been his thing since it was just a blog about apple product speculation and update. It's always been tedious.
Taking you at your word, though, I'd be interested to see what you think of the watermarking technology in a blind A/B test.
Do they also do it with code? Do you think deliberately picking tokens that are not the highest probability in code is acceptable for the consumer?
I'm skeptical that such a thing as a universally optimal translation exists in cases beyond the trivial. But if it does, I see no reason to think LLMs are anywhere close to it, so I think nobody will be able to tell the difference with watermarking.
That's certainly true for code. LLM code is at best mediocre. There is oceans of room to subtly watermark generated code without practical impact.
I think in this case it doesn't help that there are multiple watermarking schemes, and the easiest for people to understand is the red/green scheme by Kirchenbauer et al. (https://arxiv.org/pdf/2301.10226), which does technically distort the logits (but I'd argue only in cases where you wouldn't notice it anyway).
I wasn't aware of this gumbel softmax scheme, it seems you're referring to https://simons.berkeley.edu/talks/scott-aaronson-ut-austin-o... ? That's really clever as it doesn't even distort the logits, basically cryptographically indistinguishable from a "real" random sample unless you have the key.
The actual scheme Claude uses seems to be neither of those two though, they say it is SynthId-text which seems to be tournament sampling based.
Indeed.
It's frankly bizarre to see the assumption to the contrary being made by someone who's been passionately blogging by hand for years, who also happens to be responsible for the notoriously vague, humanistic, DWIMmy Markdown standard.
Prove it, then? It's not a claim that GumbelSoft paper makes: "Regarding generation quality (perplexity), GumbelSoft shows relatively low perplexity"
Sometimes when you're trying to write something, it really seems like the exact words matter a lot. Suggestions made to be more direct or use a more common word here or whatever seem to really impact the thought that you're trying to communicate.
Certainly we've all had times when trying to communicate clearly when the specific words seem very important.
Can you give an example of something smart John Gruber has said or written? Because I can't think of one, but I can think of many dumb ones.
This is not entirely accurate. Sure, there's never a token with 100% certainty, but there are often tokens with 99.9% probability, but this technique of course does not change how such a token is sampled.
Gruber made an effort to but doesn't fully understand how SynthID works. LLMs select the next word randomly from a set probability distribution, so there is no "best choice" unless you run the LLM with a temperature of 0 which would give out terrible results. Anthropic runs a non-distorting version of SynthID that doesn't change the probabilities of the underlying distribution of tokens. It makes the watermark less likely to work over smaller samples but preserves text quality. I encourage the mathematically inclined to read the paper:
Suppose there actually is a best word choice. The LLM doesn't know what it is but makes a guess. Maybe it's the best one, maybe it isn't. The probability that SynthID changes the best choice to a worse one is equal to the probability that it changes a worse choice to the best one.
> The probability that SynthID changes the best choice to a worse one is equal to the probability that it changes a worse choice to the best one.
I'm not understanding how the word with the highest probability isn't the "best choice"?
Right, I've done this, and this makes sense to me, but I'm not following how that falsifies the top probability word being the best choice in any particular instance.
"Picking only the best word at each decision point results in a worse final result" seems like an imminently reasonable hypothesis.
Does the author think he is currently getting T=0 output from Claude? Is he under the impression that T=0 produces the "best" writing?
This entire article just seems so detached from the basics of how LLMs work.
Seems like he really likes to use LLMs and is worried that quality will be degraded. But he will never demonstrate such degradation scientifically, we don’t have anecdotes even.
That's ... even worse? So we're all here in the comments trying to figure out what the author means, and what their overall point is, while clearly they don't even use the damn thing? Oof... What a waste of time for everyone involved.
It seems fully logical to me that someone who writes for a living (who, as it happens, developed the very markup language LLMs use for everything) should be invested in understanding the automatic plagiarism and word calculating machine from an intellectually honest position.
I personally am pretty severely big-two-AI-firms, increasingly anti-big-tech, but I am learning and researching uses of LLMs because for myself I really need to understand how to use them in an intellectually and (as far as is possible) ethically sound way. Learning because as a boring old freelance programmer I have to; foolish to pretend otherwise.
So I completely understand his position — that the AI industry is hot air and crooked and scammy and weird, and some of the people involved genuinely rather dark-sided, but the technology exists and if it hints at threatening your livelihood, you need to understand it.
From reading his work for the best part of twenty years or so (and emailing him intermittently over that time) it would seem to me that he's a lot less bearish on the tech industry than me, and a lot less fond of the EU than I am; he's more optimistic than I am. But he writes because he has to write. I should think that would make him highly invested in understanding what LLMs do.
I suggested he as a tech writer has reason to be invested in understanding how they work. It's not really a sustainable position to not understand, is it?
I don't think it's weird to ask that people commenting on x doing y with tech z at least try tech z, no?
Imagine this pamphlet: Basting a steak in a cast iron skillet with butter and herbs is a perversion of grilling a steak on charcoal. Signed, a life-time vegan who hasn't cooked a steak in their life.
Then imagine people jumping in the comments to discuss. Isn't it a waste of time?
On what specific basis do you assume he hasn't tried it? He's definitely blogged about the desktop apps, after all.
Or are you arguing that a writer doesn't have a meaningful or valid opinion on LLM-generated writing until they have tried to pass some off as their own?
This just seems weird to me. I mean, I have an opinion on this and I am personally never going to use an LLM to do published writing. On an intellectual level I can still see that there is nuance in it for others (for once I agree with him about an EU regulation).
But one of the issues with HN threads is that you can sometimes lose the sense of what you're replying to by clicking further down the thread, and I have committed worse misunderstandings than this, so absolutely no need to apologise (and I probably need to consider this when I am replying) :-)
His first take on this situation was cutely naive, thinking they were going to inject secret hidden unicode characters. But ultimately he has a massive hate on for the EU -- they were mean to Apple once -- and it comes out in any topic that overlaps.
He notes in various other posts that he uses AI/LLMs and chatbots quite extensively. (I don't recall what for exactly, but not for writing his pieces.)
No and no. I am not sure I agree with his point but I know he is not ill-informed on either of these points, because I mentioned them to him a couple of days ago.
The point I made (quite briefly) is that watermarking is only feasible because for good writing it is necessary to use T>0, or the writing will never explore a more creative choice, and that at T=0 you don’t even need a watermark to spot LLM-generated text.
The point he is making is consistent with this, isn’t it? Either you allow temperature to drive creativity, consistently in a way that can be influenced and analysed, or you adulterate that process for the purposes of meeting a corporate/legal directive, in a way that is proprietary and obscure. These are ethically distinct approaches, and since he disagrees with the EU objective he comes down on one side I guess.
Me, I don’t care about the hypothetical enough.
Not least because I think Claude writes depressingly badly and I doubt any steganographic change will enrage me less.
You are projecting that onto me, and I cannot tell you how comically poorly aimed it is.
He repeatedly states that choosing "the best word" is the most important thing to him. I don't know how you reconcile that with creativity itself, let alone probabilistic sampling.
That doesn't actually give you the 'best' text in any human sense of the word. Just like playing the 'best' move in Poker without sampling leads you to lose a lot of money.
This framing does not make sense to me. What do you mean by "influenced and analysed"? How have you or anyone been influencing or analyzing the randomness behind the sampling process to create better writing? What makes the unadulterated randomness "driving creativity" but a different random choice uncreative?
You're mischaracterising or misunderstanding my point, or I mangled it.
I mean it is possible to analyse, control, monitor, study the impact of changing temperature on the writing, yes?
The point about watermarking is that this relationship — change the temperature, see the effect — is now being adjusted by an unstated, secret process you explicitly can't control.
(I gather Anthropic have recently taken away this setting anyway; that was news to me.)
I think a good intuition here is that watermarking is sort of like picking a specific PRNG seed. It's not changing or interfering with the temperature - we're still sampling from the model's probability distribution. But we're making it so the analog of the PRNG seed is coupled to the previous context.
Describing it in terms of a word-by-word choice is simpler, but writing quality is dependent on the interplay between words.
"The weather today was cold and {grey,overcast}." If the next sentence is "I miss yesterday, when it was {bright,sunny}." then the choice between "grey" and "overcast" is no longer neutral. "grey" and "bright" pair together, as do "overcast" and "sunny". Or if you disagree with my aesthetic sensibilities, consider:
The weather today was cold and {grey,gray}. The {color,colour} of the sky matched my {humorless,humourless} mood.Well, not that weird actually. He just has a hard-on against anything that comes from the EU since Apple got in trouble. If the EU said tomorrow that they want peace in the world he’d be in Fox News the next day calling for an invasion. As a former reader of Daring Fireball, it’s just sad to see.
I dunno, I guess that's what you should expect from Gruber but these EU-bashing articles lowered the enjoyment I got from his blog underneath the bar for me.
After all, if it wasn't for him who would ever speak up for the trillion dollar corporation? Without brave heros like him, these poor vulnerable corporations would be facing all sorts of attacks from evil regulatory organizations and their dastardly scientists. Just think of the profit - all those euros - that may be lost like chaff in the wind.
Further he later compares Gemini to Anthropic models, saying the latter "writes better", emptily ascribing this to the synthid stuff. I think he heard that Anthropic currently has superior models, but it certainly isn't because they "write better", and if anything Opus 5 now is virtually unintelligible, before the fingerprinting.
The fingerprinting stuff sounds weird. If the EU wants it, it should be limited to the EU, and Anthropic is fully capable of doing that but clearly saw value in recognizing their own output. Is it going to destroy the quality of the output? We'll have to see, and this anti-EU piece, predicated on utter ignorance of the field, is not convincing.
It's like some weird k-pop stan sending death threats to someone that dissed their favourite singer. Just super strange stuff.
I think that was intended, yes.
A point of confusion for me, however: is every watermark unique? Is every algorithm for watermarking going to vary amongst models and amongst model versions? Will each model publisher keep this watermarking as a trade secret, that they alone can detect? If so, this can't scale! How do you detect "JoeBob 4.3 LLM" output? By querying every single model's watermark-detector? And if they all work by re-running the model and using tokens anew? That is extraordinarily wasteful.
If a watermark is not self-evident, or universally detectable, then it is no good. Take, for example, US currency. The security measures are published and well known. Any count-out room in retail has a big poster indicating how you can detect authentic US bills. Nobody has to accept non-US currency in the US, and so the only authenticity you need to worry about is your US bills alone. LLM watermarking has none of this in common. Currently sounding like a shitshow, if you ask me.
Working backwards: if it is possible to confirm 100% confidence that a chunk of text is LLM output, then it is "PD until proven otherwise". How can a human reliably assert human authorship of their source text? When all watermark tests fail? Is that proof of humanity now?
If a human proves human authorship, and LLM watermarking tests positive, then is that going to be considered a "derivative work" or not? What if there is an applicable license for the source work, such as "CC-BY-ND" that prohibits derivative works?
This has not been court-tested, and I expect that it will need testing at that level before we can have any assurances.
As for it being "uncopyrightable" if it were the output of an LLM, I think computer people are making a very aspie interpretation of a single decision. I think it's more that the LLM (and thereby its owners) cannot itself hold a copyright on its output, that output has to be touched by a person before it is copyrightable. A particular view from the top of a mountain can't be copyrighted, for example, but a photograph of that view can be.
I'm not sure it at all precludes a "robosigning"* sort of situation, where machines generate output, hired temps sign and claim that output, and immediately sign it over to the people who hired them (as a work-for-hire.) Copyright is stupid, artificial law, not logical.
-----
* https://www.mortgageauditsonline.com/what-are-robo-signers/
For the watermark to be detectable, the text needs to be like 75% AI generated.
If you have an LLM “touch” one section of the article, it’s not gonna be detectable.
Here is some text that is copyright to me. As you infringed my copyright, please pay my $5000 license fee for every user who has read it:
> Copyright law was updated in a very helpful way in the last twenty years sometime so that as soon as you post something to the internet you have copyright. If you need a citation don't hesitate to ask someone else.
Wouldn’t having that be enough to eventually reverse engineer the key?
Removal may come down to changing every third token to a different one.
And if you copy-paste the answers from LLM, I think it's only fair the end result gets flagged. You're not writing it yourself.
I wonder if it would even get flagged in that case, because wouldn't the probability distribution of a token when the LLM is suggesting an edit to your writing be different than the distribution of that token once it is in the context of the text it's editing?
It's still possible to use Claude to proofread - highlight grammatical, flow, structure, logic errors and make simple suggestions for you to pick and choose or adapt as you wish. No watermarking will flag your text. No flaw accusations of LLM authorship will haunt you. All will be fine.
But if you want an LLM to rewrite your text, that's (a) not proofreading, and (b) should be flagged as LLM generated ... because it is.
You could use your own hypothetical house elf to do it for you, or pay someone to do it. LLMs are just cheaper for a certain set of problems.
People will find ways to circumvent this, so this limitation will only hit the technically less adept people.
In the fact that you didn't write it.
> You could use your own hypothetical house elf to do it for you, or pay someone to do it.
Yes, and those would be similarly problematic (and more expensive).
This is a fact and this is generally not a problem. Customer support guy Joe did not write that email to you with a refund: someone else did it and Joe did pick the template. Alice did not write that post card to Bob, someone else did and she just googled some nice text. We deal with a lot of content that wasn’t written by the person who signed it. That content, when written by LLM, may indeed contain watermarks and nobody will care about the choice of words, because only the meaning matters in such communications.
People pay too much attention to authenticity here, which is no more than a demonstration of an effort. LLM can and should write scientific articles because the real effort is in directing research, not summarizing it. LLMs can and should write news, because it is cheap and efficient, and real reporting is in discovery. LLM can and should write fiction and make movies, because there is no reason why creators of various junk should earn their money easily. LLMs do not replace real talent. They just emphasize for an average person how easily replaceable they are. And that‘s ok. Creative industry is a blue collar job now.
In case of the former, there is a problem worth solving.
Factory farming also makes meat cheaper than organic practices. I'd just like to know which one I'm getting.
"When Claude proofreads text written by a person, what it gives back has generally only been lightly edited; because nearly all the words are the person’s, there’s very little (if anything) for the watermark to attach to. Depending on the length of the text and how heavily Claude has edited it, those changes might not be enough to make Claude’s involvement detectable."
(emphasis added)
A proofreader, like an english teacher, can return you your writing simply annotated and marked up, with suggestions and edits in red pen for example.
Then you rewrite your draft into a final using those edits and notes as suggestions.
Making an LLM act like that proof reader would likely not cause your output text to be labeled as generated, even by this system.
They will need to dodge around the EU requirements but it will probably just come down to an alternative method to watermark or a contractual assurance you won't mis-represent the source of the text.
I hope the other providers will add a geographical limitation on this EU rule.
(On a side note, I wish they would replace those EU beauracts with LLms).
but proof reading is a linter, not a writer. the proof reader will say "I think this is clumsy can you try x,y & z"
Then why are you using an LLM to write? They're not capable of understanding such nuance. They do pick randomly between two synonymous phrases, they do not use some super smart algorithm to pick the one that sounds the best.
This excuse doesn't hold any water at all - Occam's razor says the author is just super annoyed that his AI writing will be identifiable as AI writing.
But even if there were such a thing as a truly "best word", for some context, what are the examples here? Mango vs pineapple? Gray vs overcast? In what case is one of these better, that AI would normally infer but would suddenly be "perverted" by SynthID? Do you think your emotional state and preferences are being evaluated if they aren't explicitly in memory? And if they are there, do you think that the generator will bypass those instructions in favor of the watermark instead of placing it somewhere you won't care? I just. Genuinely don't get it. There may be words that matter in specific contexts or to you as a reader, so you should bloody well put them there.
Neither of these is "better" or "more precise"; in fact, LLMs will generally choose randomly between these candidates based on temperature, and SynthID should not distort the output of an LLM any more than the default temperature settings do already.
I agree that those are not the same sentences, but if the difference matters to you, you shouldn't be using an LLM. This difference exists at the level of what sounds better and is more evocative; to an LLM, nothing sounds like or evokes anything. They simply do not write good prose.
https://www.anthropic.com/news/claude-text-watermark#:~:text...
I see a section "How to install apps from a developer’s website in the European Union". Do you see that?
(They do require notarisation, but not review/approval).
It is a bit of a mystery to say that “its okay to choose different tokens that we would have for watermarking bc people don’t notice” as though word choice doesn’t matter. If it doesn’t matter, doesn’t that mean that intelligence is more of a commodity than they would want it to be?
This. I want the model I'm paying for to be "pure". I don't Anthropic or anyone else messing around with it, especially not for idiotic reasons like facillitating AI stigmatization. The "safety" nonsense is obnoxious enough.
They should train the best possible model and let the weights speak for themselves, not degrade it into some perverted form to appease people who hate AI anyway.
From the people I’ve talked to at universities, LLM based cheating in education is an unstoppable nightmare. I don’t have a problem with LLMs. But I do want the cheating to - somehow - stop. The people who cheat miss out on learning. And the people who don’t cheat have their degrees devalued by those who do cheat.
They aren't there to learn. They are there to jump through hoops to get a degree that will let them get a job so they can make money and prosper. The learning is entirely secondary.
The cheating will stop when there is no longer any economic incentive to be there in the first place. People with "pure" motives will refuse to cheat on their own, precisely because they want to learn and cheating prevents them from learning.
Watermarking AI output is a treatment for symptoms. The cause is the higher education meme. Somehow, getting a degree just became the default. Can't get a good job without a degree. That meme needs to die, and higher education will never recover its integrity until it does.
You can't just destroy some signal and handwave that you'll make it up in some other way.
There is none. It's just a jobs program fueled by student loans. Higher education in the west has been corrupt for quite a while now. AI is just the final nail in its coffin.
I agree that higher education shouldn’t be as required to get a good job. But reality isn't black or white. Reducing the entire sector to a corrupt degree mill throws the baby out with the bath water.
I’ve worked with plenty of smart, self taught programmers throughout my career. The highest paid guy I know didn’t finish high school.
Bureau of Labor Statistics. Jobs requiring higher education pay roughly 2x more than those requiring high school education and roughly 3x more than those requiring no education at all.
And even if there's no formal requirement for a degree, it doesn't automatically mean people lacking degrees will get hired either.
Anyone who wants to get a well paying white collar job pretty much needs a degree.
> I’ve worked with plenty of smart, self taught programmers throughout my career.
And how did they get the job? Networking?
> The highest paid guy I know didn’t finish high school.
I think this ought to be the rule, not the exception.
This isn’t really evidence either way. Why do companies pay twice as much for people with higher education? We can’t tell from that statistic. Maybe it’s what you learn in class that makes you twice as valuable to potential employers.
> how did they get the job? Networking?
Probably. After all, that’s how most people in our industry find work. Degree or not.
I claimed job prospects is the biggest reason why people spend years educating themselves. Evidence of 2-3x higher salaries directly supports that by providing an excellent motive.
Or the degree is something else. A marker of status. Or a signal of IQ and conscientiousness, since companies are legally barred from directly issuing IQ tests.
An observation of 2-3x higher salaries for graduates doesn't differentiate these two theories.
But rereading your comments, maybe that was never your argument. You said:
> The learning is entirely secondary. The cheating will stop when there is no longer any economic incentive to be there in the first place.
Companies might be entirely rational in offering 2-3x higher salaries to people with degrees. If that is the case, it's not just a meme. And the economic incentive will remain indefinitely. And universities and companies aren't doing anything wrong here.
Throughout your career, on the job performance matters a lot more in aggregate than your CV. Most people are hired from referrals, after all. If students value the degree over the education itself, they're cheating themselves out of all the benefits - economic and otherwise - that education can provide them throughout their career.
I don't think the models are pure in any meaningful sense. The labs have some idea of what kind of output they want from the models and then they put a huge amount of effort into training the models on the right sorts of data and massaging the models afterwards to push them towards the desired output. Then at a more practical level there's the layers of filters before your prompt even hits the model (e.g. anthropic's auto-mode classifier), system prompts, response level filtering etc.
I’d sacrifice a few %, easily. Maybe 10%. The models are getting smarter at such a fast rate that I’d be willing to lose a month or two of progress to help slow down the AI cheating epidemic.
It sounds like you expect this fingerprinting approach would dramatically reduce the intelligence of their models. But I’m sure anthropic has measured it. I doubt they would have rolled this out if the intelligence cost were that significant. I suspect the cost is less than 5%. I personally can’t tell any difference from before they added fingerprinting. I bet you can’t either.
Is the cheating epidemic so bad? I’m a little out of the loop there truthfully, what are the consequences of not being able to detect AI generated text in non academic settings? And in academic settings… maybe I am underestimating the challenge, but it does feel like the assignment and ways education happens needs to change?
As an aside, I’m not totally sure why this solution feels so icky to me. There’s something Orwellian about how the phrasing of a passage embeds hidden information that only Anthropic can see i guess
How effective is this on the new fingerprinting mechanisms?
> Is the cheating epidemic so bad?
From what I’ve heard, yeah it’s out of control. And all the existing llm detectors that academics use have a high false positive rate, which catches a bunch of innocent students in the cross fire.
> it does feel like the assignment and ways education happens needs to change?
Why? Was there something fundamentally wrong with how universities teach and assess?
The sector is responding. For example by moving back to more in person exams and reducing the load of any take home exams. Is that good, for some reason?
> There’s something Orwellian about how the phrasing of a passage embeds hidden information
Interesting. I don’t have the same response. LLMs give me an acute sense of existential dread each time their capabilities improve. But fingerprinting doesn’t move me at all. Do some soul searching on why this bothers you. I’d love to hear why, and I bet you aren’t alone.
Such as?
But it doesn't! The distribution doesn't change at all. The only thing that changes is that sampling of that distribution becomes deterministic as per a precomputed seed.
(If you don't like calling it a distribution when it's at 100% for the chosen token and 0% for all others, then look at it as an output distribution across all possible prompt inputs, or perhaps just the cluster of prompts that achieve whatever you're trying to accomplish.)
I think (or I hope, anyway) that this overestimates how much impact the tweaks actually have on text.
The model has some things it "wants" to say. If it wants to tell a story about how someone reacted to dreary weather, it's going to tell approximately the same story regardless of whether the dice-roll caused it to describe the weather as "gray" or "overcast". And because "gray" and "overcast" were _already_ possibilities, the tweak from 52% -> 55% is completely lost in the noise.
But it's true that this is all based on hope. I'm confident that you could make the tweak against arbitrary prose and even a true artiste like Gruber would never be able to tell the difference. I'm less confident that there isn't some edge case somewhere that causes a tweak to be worse than 3%, especially in some narrow application where word choice _does_ matter (like law). Even then, though, laws are already written by people who are as noisy if not noisier than LLMs.
This seems like a fundamental misunderstanding of how this sort of watermarking works. (Either that, or I have a fundamental misunderstanding of how it works lol.) It doesn't change the probability distribution of the next token at all. If you were getting XYZ 48% of the time before, you're still getting XYZ 48% of the time. What's changed is where the random numbers come from. But as far as you're concerned, there's just as random as they were before, just like an encrypted message is indistinguishable from random bytes if you don't know the key.
It is definitely blurrier whether you can say this approach changes the distribution then. By definition, it _has_ to change the probabilities of output tokens, but it's not totally clear that the pseudorandomly generated scoring functions does affect the learned distribution.
put another way, I think it's safer to do:
compute distribution -> sample -> watermark from sampled options
than it would be to do:
compute distribution -> watermark distribution -> sample
As described in the paper, you're right that it doesn't affect the main sampling technique, but what they do is they sample the distribution for 2^m samples, and then use Tournament sampling to choose the tokens among those 2^m samples, and the watermark key changes the scoring of the tournament options, using the watermark key as an input to the random generator that generates the scoring functions.
Then, to calculate the watermark, they take the text, and compute the mean g-values of the text, and a higher score means that it's more likely that it was sampled using the provided selection of tournament watermarking functions.
let's say you had some top P words: mango, banana, pineapple, guava, and you sampled 8 times, and got each one twice in the following order:
1. mango 2. banana 3. pineapple 4. guava 5. mango 6. banana 7. pineapple 8. guava
without tournament sampling, you'd truly see any of those come through. But in tournament sampling, you take those 8 options, create m scoring functions based on the pseudorandom generator, and score the 'tournament' by sampling the biased distribution you create from the g values. That does change the sampling from based purely on the LLM and entropy, but i mean, if the watermark key is also generated from some entropy, it's probably representative of the original sampling options as expected?
this is a very fascinating topic! I do still stand by my point that anthropic is the only one who can tell if something is watermarked or not and feeling icky, but the paper has mostly quelled my concern on impacting the intelligence part.
Pasted below:
Finally convinced myself that non-distorting watermarking is real, a la Anthropic / Google SynthID. It's not just spin / marketing. This really is a "Monty Hall" like problem. (h/t @random_walker for that analogy.)
Sharing here in case it helps anyone else.
On one hand, watermarking is "obviously" damaging to the output. The LLM does all this work to compute token probabilities... then you essentially perturb the probabilities? Of course that's bad! Every single token is using perturbed probabilities!
On the other hand, you obviously can freely inject a layer of randomness into sampling. Take any categorical distribution. For each sample, perturb the probabilities, then draw the sample. Done correctly, over many samples the results will match the original distribution.
It still kinda feels like magic, but it's not surprising that you can take that core trick and shape it into a non-distortionary watermarking scheme.
(i.e. it gives you sequences that were just as likely to be generated by the original model as any other sequence)
I honestly can't stand the way Claude writes. This watermark change just makes it scarier.
No it's not. The bad part about it is that some machine is writing instead of you, not the specific stylistic idiosyncracies.
Well, akshwally...
> Interoperability. Providers must implement an interoperability solution for watermark detection such as a standardized API access method, a publicly readable signpost mechanism embedded in content, or participation in a consortium detection solution by February 2, 2027
LLMs have never been the place I've thought to expect any commitment to the craft of writing, to be fair.
Oh! If you want that, you should run your own model and set the generation temperature to 0 :) Because that's not what any commercial LLM is doing. Never has been. This is just making up a universe that doesn't exist so you can get mad about no longer being in the universe that doesn't exist.
The masking technique of using a subset of the statistical distribution for each next token isn't going to be meaningfully distinguishable from a natural language perspective. I honestly think its a very elegant way to implement watermarking. I've got no real opinions on how effective it will be to people actively trying to defeat it, but I suspect that the people who are trying to pretend that LLM text was something they wrote themselves are probably too lazy to put in the work to try and defeat it anyway.
At the point that you're generating entire volumes of text from Claude you're not really trying to be a sophisticated writer. I don't see how it's going to hurt for it to choose random related words.
Alternatively, you ask Claude to write the whole thing and proof read it yourself. In that case I'd like to know how much you'd need to change to break the watermarking, i.e. how much of a text would you need to change for it to be considered your work and not that of Claude?
Who could have seen this coming???
If this didn't have a detectable effect on the quality of the token probability calculation, the watermark wouldn't be detectable. It may be a small degradation in the quality of the output relative to the neds of many users in many situations, but it's not zero. It's literally sometimes choosing different words that it otherwise would specifically for watermarking purposes.
If this was not the case, the encryption would be broken, and most everyone agrees that good encryption does exist.
The "quality of the probability calculations" as you put it is 100% in this case and any less would be a huge deal (as in - breaks all of the internet).
So, now you just take those same random bytes and use them as the seed for your LLM token choices. The output has the _cryptographically_ proven exact same quality as if you were using a true RNG (which you likely weren't using anyways).
You just need to know your LLM distribution and the encryption key, then with each new token you exponentially increase the chance of knowing whether it fits your encryption key. Without actually affecting the token choice in a perceivable manner.
> choosing different words that it otherwise would
The "that is otherwise would" is carrying all the weight here. "Otherwise" is sampling from a distribution. You just sample from the same distribution but with a cryptographically secure, seeded RNG. https://en.wikipedia.org/wiki/Cryptographically_secure_pseud...
"Knowing the LLM distribution" seems to me like the only hard part because you don't know the context of any random snippet.
I struggle to understand the relevance of that comment.
The blue/green token list biasing process literally does cause different tokens to be occasionally chosen. Not only that, but because a different token was chosen at one point, this changes the probabilities of all subsequent tokens, and the resulting later token stream every time it happens. If you had access to the token stream as it would have been, and the watermarked one by the end of the text they will be very noticeably different.
So, you don't own the generated text, and can't use it freely then. What if I copy paste a section, or rewrite a section of text to my liking? What if I rewrite some lines of code that contains the mark?
Security theater, and vague enough to be used as a weapon against who the government wishes.
I hope it's left off for non-EU customers.
Try running an llm like qwen 3.8 27B in Q8 locally with an intentionally very low temperature setting, it will write like a caveman crossed with a robot. You may find that an extremely literal output does not look pleasant to read for humans.
"I've been working in the media and model IP space for quite many years. What this article misses a bit is the threat model for watermarking in general. Watermarking and fingerprinting have inherently weak security guarantees -- they rely a lot on security through obscurity, weak assumed adversaries to deliver. There are clear trade offs between true positives, false positives and maintaining the quality of the media. It's true for audio-visual media, models and their outputs alike. As much as I like to take shots at poor technical choices by corps and govs, this one is unjustified. Sure, inserting glyphs is bad but biased sampling is as good as it gets in 2026."
Since the announcement, there have been many people who don't seem to fully understand what a security guarantee is, what trade offs it might involve, or how popular the type of technological solution is in general (media watermarking is ubiquitous).
And naturally, there are challenges with how you will make sense of the score in your org, e.g. you wrote an email, and it's flagged as LLM-generated because you copied two generated/edited paragraphs.
Yeah, the article might disagree with watermarking as a matter of principle, comparing it to censorship. But the methodological arguments that I have read so far have been thin in these articles.
The thing I don’t understand is why he seems to care so damned much about this subject--enough to write over 4,500 words on it!
John writes for a living. That’s his profession. He’s been writing for over 25 years now. When you’re that good at writing, and you care this much about your writing, you don’t allow an LLM to take over your job. I just can’t imagine that he’s in the market for LLMs and that literary excellence is his number one selection criterion.
So why is he so livid about it? It’s like being angry that wine is going to start coming in smaller bottles even though you don’t drink wine.
Even if he’s angry on behalf of other people, I don’t get it either. In my view, having LLMs write publishable content on your behalf is not a socially-acceptable use case, nor a professionally-acceptable one in most professions, even though people are abusing it for this purpose anyway. And besides, the models aren’t even all that good at it today. If you agree with that, then you certainly should not care if it’s using different phrasing than you otherwise might prefer if the meaning is similar enough.
I can't help but wonder if perhaps his hatred of EU technology regulation (which, admittedly, is mostly pretty dumb and is mainly just making life worse for users) is getting the better of him.
Has that ever been the case? Are they not actively tweaking their models, their fine tuning, the system prompts, the tool definitions and implementations, the guard rails, tool calls, instant responses. There are hundreds of knobs that they can change daily, or between each prompt, or even half way through a generation.
> At each decision point, they’re a little more likely to pick a word from the green list than the red list.
Wrong. There is no global red and green list. It's dependent on context and balances out on average. It won't change the result when one token is predicted overwhelmingly likely.
If Gruber can't tell a fully AI-generated article from a human-written one, perhaps he shouldn't care so much.
NB: I was told yesterday it's apparently a meme to call out Claude-generated output, but here I am, as I believe it's quite relevant to the topic at hand.
But it feels to me like you would need a hell of a lot of text to bury even a simple account ID. The nudges they are talking about are of the order of a handful of bits over several hundred words, I think?
Reminds me of printer tracking dots.
It's a good idea for many human endeavors to be able to identify AI writing. Communication, after all, is our main way of building the social fabric.
However - and crucially - good writing is still beyond the frontier of any model I've seen so far.
Watermarks for the things that truly matter may not be important at all.
Finally, as X commentators have shown, simply removing punctuation or changing a word here or adding an adverb there manually will screw up the whole process enormously.
The best will be the clever folks who retroactively apply the model distribution to fraud or other crimes to try to implicate the companies via watermark.
Gotta feel for their product, policy and legal team.
"Just flip a coin to pick a random synonym. Who cares?" - AI Labs
“Absurdly and insultingly”? Come on…
If all the providers use watermarking systems with different shifting logit weightings, and the keys are secret, you have to check every provider to see if it produced a given text.
Which is clearly ridiculous.
And if all providers collaborate and use the same weightings, or if the weightings are constant and not rotated cryptographically, a generic watermark remover becomes trivial.
That's not even getting into the legal complexities of businesses running open source models without watermarking locally.
Either they false positve on pretty much everything ever written, or the chances of catching a true positive is so low as to be useless.
Basically Cinavia for text, and that often falls over and is easy to remove even when there is megabytes of data streaming over a long period of time rather than 2 or 3 bits per wall of text, let alone what most people use claude for, when there is a strict dictionary and other tight output constraints.
Everyone who invests in AI companies wants to see the value of their investment increase.
I'd give it about 3-5 years until an AI company claims copyright over code their LLM produces. This is a crucial step in that path.
But if I craft a prompt that doesn't leave space for creativity, how can they include a watermark? e.g. "Rewrite the following text, replacing 'foo' with 'bar'."
I'm curious to see where they draw the line, and whether the watermarking really affects the (perceived) quality of results.
I was bulding a small interpreter and writing an article in ~markdown yesterday with Fable. And while it codes like a pro, it writes like a sixth grader.
Let's see how these watermarking stats hold up if/when llms start writing well.
The author has expressed a preference. Assume that there is a sequence of tokens, such that it is considered the absolute best by the author. This particular method of watermarking makes it less likely to generate that sequence, by definition.
I feel their argument would have been clearer and stronger if they had spent more time exploring the alternatives, and whether these alternatives would be just as effective. It is trivially easy to remove invisible tokens.
Like it or not, there is a public good to being able to identify AI generated content, and a small degredation in quality is tolerable in my opinion.
I don't think anybody has to worry about this issue though. Manual writing, coding, and proof reading continues to be an option. Where AI output is nothing to be ashamed of, the tools are available. For everyone else, there will be LLM providers that ignore EU law.
You can't assume that because if that was the case he'd already know what sentence to write, because that's what that means.
The notion of a best sentence requires a final cause, an end to write to. By their very nature that's not how LLMs work, so you can't 'degrade' them on that front. They can't lose a property they didn't have.
* Shows the real result from their detector instead of manipulating them, you have no way to verify * Has good enough security to prevent a key leak * Rotates keys to reduce the impact of a leak (once a key is leaked anyone can rewrite text to look more/less claude generated and it becomes useless) * Will not secretly give watermark-less access to governments or high profile corporate users * Will not use multiple secret keys to track individual users. This one might be less realistic because embedding ~32 bits of signal would probably affect quality a lot more than 1 bit.
And don't forget that the detection API will work as an oracle. If it detects your content you can send it to a different model and try again until it comes back clean.
I wonder if this is why Opus 5 keeps writing excessively long comments, even though I keep instructing it not to (both in chat, CLAUDE.md, and in its memories)
If the same red/green algorithm described in the article is applied to generated code, I cannot imagine how that does not degrade code quality (probabilistically, not at every point).
The part I agree with:
It is true that watermarking can be done by "just" swapping one PRNG for another, and it is even true that with today's LLMs, it is possible that this will not degrade the output. But it has a cost, and as things improve, that cost will matter. You are intentionally reducing the degrees of freedom in the output, and using those bits of entropy for a purpose that does not improve the quality. If you maximize your tradeoff of bits for quality, those extra bits lower the ceiling of what's possible. It's a very simple information theoretic argument, and the only plausible argument against it (that we're using those bits so inefficiently now that the new PRNG is no worse than the old) only holds in the short term.
I also agree that having TOS that forbid removing the watermarking is deeply, deeply problematic. Hell, the whole essay is well-written and persuasive, and gives good reasons why this is a poor approach.
What I disagree with, and the reason for this comment, is the entitlement.
> The idea that anything other than my needs should factor into the generation of text for me is patently offensive.
This attitude is what is patently offensive for me. This is the argument that the world is beholden to my interests. It says that worrying about negative externalities is immoral. It's another form of certain people being above the law, shareholder profit maximization über alles, might makes right, we have to do it or someone else will, "we just help people connect", {code,a gun,roofie} is just a tool.
So I agree that the watermarking has a cost. But you can't leave out that it is an attempt to reduce negative externalities of AI. Whether it's a realistic or worthwhile attempt is a whole other debate (and Gruber does a good job of debating just that in the latter part of the essay), but saying that the user's needs are the only thing that should ever be considered is reprehensible.
Computers are tools that exist to serve. Creating some bizarro process where we are compromising the technology in service of it's owner to achieve some nebulous goal is gross.
Anthropic is crowing about this achievement because they are afraid of the dirt cheap AI models coming out of China and eventually other places impacting their valuation. Full stop. There's some vague notion of preventing harm without any backing, but a very real cost for startups to develop a compliant watermarked AI model.
> Anthropic is crowing about this achievement because they are afraid of the dirt cheap AI models coming out of China and eventually other places impacting their valuation. Full stop.
It's the "full stop" that I'm disagreeing with. Yes, you can make valid arguments about the motivations behind this, or the effectiveness of it, or whatever. You can even conclude that it's a net negative. That's my current leaning. But "computers are tools that exist to serve" is an excuse, a conscious decision to abandon responsibility. Hammers, nails, social media, biological weapons, and date rape drugs are tools that exist to serve.
[1] Ok, fine, at least one started out as a sleazy way to talk about people's physical attractiveness and behavior anonymously with repercussions, in a way where the victims would have no meaningful recourse. For the sake of my argument, pretend we're talking about Friendster instead of Facebook, please?
I love your point about social media. IMO, initially they didn't know what they had with social media - it took the right kind of uninhibited sociopath to monetize/weaponize it. That's why the MySpace guy is traveling and taking photos while Zuck is building island lairs.
This is where a functional government would step in. There's no magic in social media technology or LLMs. I'd think of them as a car. You can buy a Nissan Leaf of a Porsche 911. One is an ok basic car. The other is an (over) engineered experience in the form of a car. You cannot legally drive a Porsche to it's potential on the public highways... because we have laws that regulate driving and hold the operator accountable.
We're thinking about Claude Code or Gemini or whatever. The people running these companies are like "I want to exceed the power of John D. Rockefeller or Stalin." If we or the EU are going to regulate AI Labs, you need to grab them by the throat and they should be screaming about it. They seem to be very pleased with themselves. Barring real regulatory teeth, we would be opening the aperature to the Chinese companies to rationalize the valuations.
I agree with you, in that the benefit of the commons (preventing damage AI is doing to the world etc) is better than one individual's 'right' to a perfect product.
But I don't think Gruber is saying quite this. He's worried about genuine semantic and intelligence loss, using the example of swapping two words:
> The semantic difference between banana and pineapple is just as noticeable to the human eye as the taste of the two are to the human tongue
Further, is this really a useful way to reduce the damage of AI? It only appears to work if someone already believes and then proactively checks if text is produced by AI. It doesn't really address the core issues.
The question is: does this make my project "AI" or does it make it mine?
People use professional editors all the time. But they agree to give up any copyright in their edits. If I approve a change made by an editor, it becomes mine. But there is this idea that approving an AI sentence still make it AI.
We've had so many advancements in LLM samplers for improved text generation (off the top of my head: min-P, adaptive-P, XTC, DRY, p-less, Top-H, Top-n-Sigma, and so many more) but hosted LLM APIs only provide three basic knobs: temperature, top-k and top-p which are old as the mountains in LLM years at this point.
One thing that doesn't help the local LLM case is that all the popular VC backed local LLM wrappers also only support the same three ancient knobs because I suppose they're more preoccupied with their next fundraise than with keeping up with the advances in tech.
Many writers use AI to edit their own words, and this watermarking poisons the well for that use.
I think the backlash from this could be the seed that undoes their attempted 2 Trillion IPO this fall.
[1] https://doc.searls.com/2026/08/17/you-can-hear-the-squeak-of...
Also this:
"> I want any LLM I use to choose the very best, most precise words at every single decision point."
I don't think the author quite appreciates the level of randomness here.
This is more subtle then Claude changing prose.
Again - I suggest that the author would have to run a test on themselves to determine if they can actually find a difference.
And what I mean by that is that companies that are at the top tend to make anti customer decisions because they have lost the concept that pleasing customers matters as priority one.
How is this supposed to work in an actual lawsuit? Will Anthropic offer some sort of tool / (paid?) webservice to check for watermarks using that "secret key", and a judge is supposed to just believe whatever that tool's verdict is? And then it takes the EU another 20 years to understand what a silly idea this was?
I see even more problems with this. To check any text for those watermarks, it needs to be sent to dozens of AI companies to check, potentially paying them all for just determining whether it matches their watermark, and more concerningly sending all that mostly human-written, often high quality text like unpublished research or books, to AI companies that almost all proved to obtain training data through all kinds of dubious ways.
The obvious conflicts here are wild.
The real argument should be watermarking itself. I don’t want my shit water marked if I ask you to just rephrase a certain part.
> "My error was believing Anthropic"
The error is rearranging part of one's life around Anthropic.
There is evidence this exists already, and it's why "I have to be honest..." and "This is the right lens, ..." keep popping up.
Probably would be cheaper too. Maybe I am missing something?
Who is even asking for this? Sounds like something some obsessive internal employees would push on the world.
And this attitude is incompatible with any models produced by frontier labs. Your needs will always be subordinate to and in service of the needs of the corporation that produced the model. And we haven't even gotten to ads yet.
Further heck: It can be said it ain't even writing.-
The premium upgrade will be "obfusicators" that further modify the text sporadically to make it look more like a human wrote it.
Students for one, will gladly pay.
It already fails. It randomly picks between close candidates. To help fool people into believing in intelligence claim, I guess.
Is Gruber only now waking up to the idea that LLM companies do not give a shit about the quality of the writing they generate? That they're disdainful of the entire concept of writing as a profession? What rock has he been under?
If you care one iota for the quality and craft of your writing, you would already recognize that any degree of AI "processing" obliterates stylistic choices. You're left with readable text, but it's fluffy, meandering, and has a cadence to the writing that screams to the reader "every second you spent on this was wasted."
No, you don't. If you wanted that, you would set the temperature parameter to 0. But that would lead to less desirable results, not better. LLMs do not set the temperature to 0; they typically set it 0.4-0.7.
Apparently superfluous descriptions, unnecessary words, and paragraph footnotes are all fine, but gumbel softmax (or whatever's going on here) isn't.
The standard procedure to do this, is to chain translations to other languages and back. The message remains, but the wording will pick up some noise. --Dec 30, 2008.
Not that it's the ideal solution. But it would be really neat.
And this is why the latest Claude models blather so much more. All for the sake of this "watermark."
I don’t see how this makes it worse.
Yeah, that about sums it up!
Somehow they'll find a way to use this for regulatory capture
How Claude's text watermarking works
A company the size of Anthropic would not voluntarily jeopardize their massive valuation if they didn’t feel the resulting output would maintain a similar level of quality as before. Is there a similar worry that their system prompt, which is injected at the start of every conversation also influences token generation in an artificial way?
If regulation will ruin Claude as a product, market forces will fill the void. There are also a ton of open weight models to choose from. It’s going to be okay.
> Also, what happens if another major global market makes it unlawful for AI to secretly watermark generated text?
> We’re applying watermarking globally at launch because we don't yet have a durable way to scope it by region. However, we will continue to evaluate different approaches, and will share updates when we have them.
So unless they figure it out, would that 'major global market' essentially need to be the US?
The phrasing of the announcement implying that phrasing and diction don't change the meaning of text is insultingly dismissive of the whole field of literature, and I can see why people might take it as an afront, but the actual technology shouldn't have a negative impact as far as I can see. It seems to me that this one is more of a PR problem than something with real-world impact.
what if the llm should
- repeat something verbatim (important in a compaction prompt)
- there is just one correct order of tokens for a somewhat long chain (a certain sequence of control signals)
- provide a diff of 2 inputs
without punctuation or whitespace wiggle room?
how does the drifting work?
does it postpone the drifting and drift stronger later?
what if max_tokens is set to a low number?
in what way does this not affect output quality?
The key to understanding the watermark technology is to realize that the model is/was already using randomness to select among the top most probable tokens, often randomly picking between choices of [nearly] equal weight. The watermarking does NOT change the distribution of the random number generation nor does it affect the range of probabilities for which tokens are being considered. Instead, it only drives the sequence of the random numbers such that they form a cryptographically generated known ordering pattern that is determined by the secret key generating the pseudo-random sequence.
As a result of the approach, assuming inference is being done with all other parameterization of token selection being the same, there should be NO impact on the output quality....the amount of variation of output is within sample of the variation in output that already exists run-to-run of the same prompt. FWIW too: Google has confirmed this experimentally as well through full scale tests and evaluation of online Gemini output in search result pages.
Total variation distance has been measured to decrease as you scale a model, and that is the primary mechanism "watermarking" as discussed in the Anthropic announcement relies on. It becomes more difficult to reliably detect text as a fixed sample count without tweaking the distribution further. Either way, it's a minor problem that will be addressed over time, compared to the issue of who can detect this without guessing or developing their own sets: providers not releasing a way to detect any such watermarks without going through them makes this entire approach hostile to the public. The EU regulation on this subject is interesting, although again most certainly not the primary driver for these practices:
"1.1.2: Signatories will ensure that AI-generated or manipulated content is marked with an imperceptible watermark, with the exception of very short text. For free-form text longer than 200 tokens, watermarking still needs to be applied, even though it may have lower reliability compared to that of watermarking very long text"
A proper, effective and useful law would have required providers to regularly release datasets to run your own verification on any text released within a fixed interval of time, presumably once out of rotation. Instead, it only talks about exposing an user interface going through their own services:
"Signatories will ensure access to their detection solution through a user interface appropriate for the audience of end-users that may eventually be exposed to the content generated or manipulated by their AI system. [...] Any restriction to the access will be limited in time until more reliable and robust detection mechanisms have emerged and have been adopted as the state of the art for detection mechanisms for the watermarking of free-form text evolves."
Most interestingly, in line with the EU's mass-surveillance program, an alternative solution to watermarking where it may not be sufficient is also suggested, although only optional for now:
"Where appropriate and taking into account potential trade-offs related to privacy and security, as well as scalability challenges and costs, Signatories may implement as an optional supplementary measure fingerprinting or logging solutions for AI-generated or manipulated content which allow for checking whether content has been generated or manipulated by their AI system. For example, direct logging may be appropriate for text content, whereas fingerprinting approaches may be preferable for audio and visual content."
Since most Americans are in the latter camp, anything that even hints at making companies responsible for anything is viewed as being squarely the former.
Whereas most EU regs are "play nice, be responsible, behave like adults. If not, this can always turn ugly". Same here.
In 2005, FFII predicted that the Data Retention directive would not pass the Courts.
It just took 10 years for the CJEU to strike down the measure, as it is "mass surveillance".
Logging everything your bot does by law is of the same dimension?
Generative AI is a perversion of writing
Variable name perversion incoming...
> One of my fundamental problem with this is that no two synonyms carry the exact same meaning. “He leaped at the chance” and “He jumped at the opportunity” are very similar sentences expressing the same general sentiment, but they are not the same. The exact words we choose when writing matter.
Doesn’t make sense at all in light of the actual approach, they’re just choosing a different RNG. It’s not like they’re corrupting it by flipping words.
Should add I don’t support the watermarking and requiring it is idiotic.
Perhaps it would be useful to publish examples of samples with/without watermark. I'd suspect that the variability from simply sampling repeated times would dwarf any semantic differences you'd detect with the watermark.
Gruber's point is to bash the EU. He couldn't care much about anything else.
Except for the input of the hundreds of stakeholders they consulted, Anthropic included [1]?
> The Union is stuck on major economic crises (electricity prices for instance) because nobody can agree on anything.
That sure seems relevant for the implementation of AI watermarking...
> However, the bureaucracy forces tech into a privacy nightmare.
No, this transparency allows consumers to more easily detect AI generated content.
[1] https://digital-strategy.ec.europa.eu/en/policies/code-pract...
I would say that's more like because the US has arranged for Europe's fossil fuel energy sources to be disrupted or cut off:
* Libya - NATO made a pig's breakfast of that, it's a failed state now.
* Iran - transitive sanctions, because why not prevent non-US states from trading with each other.
* Russia (& Kazahkhstan) - The US (with or without Ukranian involvement) bombed the NordStream pipeline(s), led the EU into the proxy war in Ukraine and a sanctions regime against Russia. Kazakh oil goes to Europe through Russia.
* Gulf states - until recently, possible but not very convenient ; since Feburary of this year, the war on Iran messed that up badly too.
the US is the winner here not just geo-politically, but also as an oil exporter, with the EU now depending on purchasing US-exported oil.
Reminds me a bit of that 'Yes Prime Minister' sketch about nuclear deterrant, when the skeptical conversant asks the PM: "So what is the last resort, Picadilli?"
Guilt with a corollary of subservience to “Western values” having become the predominant ideology in Germany, strategic shortsightedness and failure to properly “relaunch” after 1990, also the occasional murder of more “promising” members of the political spectrum (sans doute at the hands of our dear “friends”), brain drain to the U.S., catastrophic “investments” like Chrysler or U.S. telcom, and barely anything of it ever raised to the threshold of public debate and intelligent reflection : a lot of things combine to explain the dismal state of affairs in modern Germany.
It cannot be ruled out that the German gov was complicit in the bombings; not wilfully, but passively, like someone too weak to resist for lack of self-esteem. Like chancellor Scholz on Feb 7, 2022, at the White House press conference where Biden threatened the pipeline:
▪ “If Russia invades, that means tanks and troops crossing the border of Ukraine, again, then there will be no longer a Nord Stream 2, we will bring an end to it.” ▫ “But how will you do that, exactly, since the project is within Germany’s control?” ▪ “We will… I promise you we will be able to do it.”
There was no reaction from Scholz. Just nothing. By the way, could be I'm wrong, but that part of the press conference seemed scripted/scheduled to me. Not the exact words, but the contents.
The trouble with drawing any sort of line is that when you're complicit and “aiding and abetting”, there's simply no motivation. And as long as the populace doesn't have a clue what's going on, their understanding being reduced to “evil Putin”, that's not even much of a problem.
"Important" refers to impact, "significant" to the statistical qualities of the underlying hypothesis, "substantial" to the work involved, and "notable" is a referential judgement by the speaker. The implied normalization of words and their respective meaning also marks one of the mechanisms how "slop" is typically creeping into the productions of "broad verbose interchange replicas" (it's all interchangeable, and a choice isn't really that, a choice, isn't it?).
Yeah, John. We're all OK with that.
I can't keep up with the current Chinese psycop. There should be some kind of status page like whywewantamericatofailataitoday.ai so we can keep up with it.
He had a very similar rant a couple of days ago when he somehow thought that they would use invisible characters. It’s just as useless.
https://www.youtube.com/live/2Kx9jbSMZqA?si=0QgCPBX2_KPZ0QTU...
> The difference between watermarked and un-watermarked text will not be distinguishable to readers
https://www.anthropic.com/news/claude-text-watermark
Which is to say, it does not actually meet the EU AI act requirements which require transparency to humans. Not to mention that if the detection requires access to the base models, it makes anthropic the only entity who gets the say on if a piece of text comes out of Claude. Anthropic is both the player and the referee here.
If there is one takeaway you should have from this fiasco it is that you should be wary of using tools that doesn't serve your needs and your needs only.
How would that work? Claude appending " written by AI" to each of its messages? That would both be impractical and useless.
If you get caught uploading watermarked media without the clear label, you're in big trouble, mister.
This is akin to adding a giant watermark on things one would made with a free product "Made with XXX". Except you're paying $200/month for it, and there's no way to disable that watermark.
I don't disagree with EU regulations, but I strongly believe the onus should be on the content publisher, not the toolmaker. If the toolmaker watermarks whatever his tool produces, it opens a giant can of worms that cannot be closed. That means anything and everything you make with this tool is no longer fully yours, it contaminates everything and makes your work traceable. Who wants that?
I was already annoyed by the fact that Claude marked everything it did on my repos under its own account (I didn't ask for any of this), but now everything is invisibly marked, even the code. Not that I care that my writing would be watermarked, since I'd rather write my stuff myself, but code? No thanks.
Meanwhile I'm running a DeepSeek V4 Flash or Pro, or a Qwen3.8, and it writes my code without a peep. Resulting repos are clean, just the way I want them. No 'Claude' account, no watermarking, nothing. I won't be looking back after having tried these new models. Whoever makes good models that don't broadcast their maker will get my business.
This watermarking will simply push people more towards Chinese models. Keep pushing in the wrong direction Anthropic. Doing this right before an IPO is a great idea.
It exhibits an undeveloped understanding of LLMs, and a righteous view that generated prose should assimilate... which should be offensive to organic intelligence.
Issues with the proprietary nature of Anthropic's watermarking aside, we will look back on this as a 'thank god' moment in the history of LLMs.
FITFY.
I have no sympathy for writers whining about what the AI is doing to 'their' writing. It's only your writing when you write it. There's any easy way to avoid this: don't fucking use it. Use you own brain.
I think they need to get their shit-together and realize this is a death warrant for the tech ( in my opinion ).
I swear, one of those days I will get into politics just to fight those two things, and the cottage industry of batshit crazy lawyers that gave birth to those things.
It's either don't track or ask for consent. The fact that the industry chooses to track is not on the EU.
Of course this misses a bigger point that tracking in the web moved in a direction that requires no cookies whatsoever, and if anything, feels more pervasive than it ever was.
And it misses the even bigger point, that the morons who legislated cookie consent did not notice either of those two realities. And the same thing is already true with the AI act; the text watermarking is trivially defeated and everyone knows it. And I'd bet it will remain a requirement for the next decade or three.
And the move away from cookies was addressed. That's why we have the GDPR now.
We can always write new regulation, but its important to handle things as they come up and not use technological change as an excuse to do nothing.
This also, just another precedent of anti-user, pro Authoritarian, from LLM companies.
You think someone will write Nazi-promoting AI policy like this when society is encouraged to look at their families as examples of neo-Nazi corruption? When their wives' and kids' friends spurn them while their families engage in obvious criminal activity to harm world productivity?
Critics need to be more precise.
I think this would be a good guard on websites to use it as another content protection layer.
The output quality will likely suffer as stated in the article although this can be mitigated to an extent by only enabling it on more irrelevant filler text while leaving the more functional sections untouched.
The solution using Unicode tricks amounts to malicious compliance as only the most unsophisticated users are going to fail to remove the AI watermarks when trying to pass off AI slop as their own prose.
The real solution here is not having stupid EU–tier laws in the first place.
Perhaps it's not actually irony, perhaps it is hypocrisy.
This isn't to say there aren't correct statements in the article, but its framed very strangely.
While I generally hate legal mandates to stab your customers in the back, in my opinion there is no harmless way to use AI and mandatory text watermarking is a good bare minimum. The EU probably made the right call. The entire AI space - open models included - is predicated upon worker exploitation, replacement, and deskilling; we should at least be able to know how much of our media diet has Anthropic's fingerprints on it.
A lot of hay is made over the pretraining process in which copious amounts of stolen data are trained on; but parallel to this is a huge data labeling and human feedback operation staffed almost entirely by people in third-world countries with robust English as a Second Language (ESL) programs. The thing is, AI models already watermark their text, they just happen to do so with the textual watermarks of the Indians and Nigerians that the AI companies hired to do RLHF because they were cheap. That's why certain AI models love the word "delve" so damned much. It's neocolonialism, designed specifically to do the kind of replacement the anti-immigrant idiots keep screaming their heads off about[1].
Furthermore, as we've seen with Hank Green, even non-cannibalism-adjacent AI usage is a recipe for worker deskilling and AI psychosis. The other half of the RLHF pipeline is to turn a pile of compressed text into a chatbot that feeds you a steady drip of unsourced information while praising you every time you spot one of its lies and never saying no[2]. This is a recipe for addicting your customers.
Also, this might just be because this is on daringfireball.net, but I can't help but think the author has an axe to grind against the EU because the EU mandated Apple sign third-party app stores. The fact that he's balking at Anthropic not going along with gating the watermarks to just the EU seems downstream of this - "why aren't you maximally attempting to resist the EU?"
[0] https://www.youtube.com/watch?v=YCPAIg7RUq8
[1] To be clear, upwards of none of the far-right have actually clued into the fact that AI is trained by underpaid immigrants, mainly because it doesn't fit the narratives the people running the far-right want to push. There are some AI robotics companies that are even very explicit that their robots are there primarily to launder foreign labor into rich companies and make permanent labor arbitrage.
[2] Continuing on from [1], the far-right actually really loves AI specifically because it rarely says no to even stupid ideas, even if it's also a manifestation of everything they claim to hate.
... and the opposite is also true, sometimes it will increase the probability of choosing the "best" word choice. So watermarking makes the LLM quality better then? /s
Or just em dashes? /s
> It’s unacceptable for a tool to sacrifice an iota of clarity, coherence, meaning, quality, etc. for the purpose of (watermarking)
And an example of the impact of watermarking on word choice [0]:
> The results of the study were quite [important || significant || substantial || notable]
The meaning of the sentence to changes slightly even in just this tiny example. Imagine the degradation when applied across an entire response!
I struggle to see how this could possibly be useful unless there's some sort of psy-op going on to trick people into uncritically accepting anything lacking a watermark as not being AI-generated.
A couple of days later it was up with the most comprehensive review of the available info summarised and referenced ready for a human researcher to explore.
This was only possible due to AI. It would have taken me weeks to chase it down and summarise it, so it would never have happened. Let's get off our high horse about AI writing.
The "watermark" can be trivially defeated, but may be enough to satisfy the letter of the law, and like many people here, I would argue that if you are letting Claude write for you, you've already accepted getting the literary equivalent of turd soup, so the harm is — or at least could be — fairly minuscule.
[1]: https://digital-strategy.ec.europa.eu/en/policies/code-pract...
(FWIW I have a more favorable view than most people seem to of the EU's efforts to at least try tackle problems like this — but predictably, the bureaucratic "solutions" they come up with don't work, but do make things objectively worse)
Watermarking is bad not just because of the principled stance that your tool should not be working against your own interests (the passionate argument in TFA), but specifically because it lends credence to the idea that AI detection is a valid and possible thing to do perfectly.
As technologists of course we know "oh well yes but with some confidence interval we can detect AI token bias across a large corpus of text." To JimBob in charge of publishing your paper or reviewing your PhD submission, all he knows is "anthropic says AI detection is possible so this 30% chance your paper was written by AI means you've plagiarized." Do you really think you're winning the argument with the certified, law-approved plagiarism detection machine? No, you're not, and your career is over.
It's irresponsible to develop watermarking because it is not anywhere close to a perfect science, but it will be treated like one by people with the power to ruin your lives. Even if you've never touched AI in your life, your paper is going through the "maybe it says you cheated" box, and you better hope those dice don't come up snake eyes.