[HN Gopher] DeepMind debuts watermarks for AI-generated text
       ___________________________________________________________________
        
       DeepMind debuts watermarks for AI-generated text
        
       Author : ambigious7777
       Score  : 29 points
       Date   : 2024-11-05 12:50 UTC (10 hours ago)
        
 (HTM) web link (spectrum.ieee.org)
 (TXT) w3m dump (spectrum.ieee.org)
        
       | ksaj wrote:
       | Some of the watermarking is really obvious. If you write song
       | lyrics in ChatGPT, watch for phrases like "come what may" and "I
       | stand tall."
       | 
       | It's not just that they are (somewhat) unusual phrases, it's that
       | ChatGPT comes up with those phrases so very often.
       | 
       | It's quite like how earlier versions always had a "However" in
       | between explanations.
        
         | GaggiX wrote:
         | ChatGPT does not have a watermark.
        
           | jgalt212 wrote:
           | I suggest we "delve" deeper int this problem.
        
           | aleph_minus_one wrote:
           | What makes you sure about that?
        
           | sunaookami wrote:
           | It has a rich tapestry of watermarks.
        
         | fkyoureadthedoc wrote:
         | Coheed and Cambria were using ChatGPT this whole damn time, smh
        
       | tokioyoyo wrote:
       | Correct me if I'm wrong, but wouldn't it simply drive people to
       | use LLMs that are not watermarking their content?
        
         | ndr wrote:
         | Agreed, that's the obvious prediction. They're also going to
         | perform worse on 3p benchmarks right?
        
         | aleph_minus_one wrote:
         | I think your idea is basically right, but there are two points
         | to consider:
         | 
         | - Your hypothesis only holds if the alternative LLM is also
         | "sufficiently good". If Gemini does not stay competitive with
         | other LLMs, Google's AI plans have a _much more serious_
         | problem.
         | 
         | - Your hypothesis assumes that many people will be capable of
         | detecting the watermarks (both of Gemini and other LLMs) so
         | that they can make a conscious choice for another LLM. But the
         | idea behind good watermarking is that it is not that easy to
         | detect.
        
         | kranner wrote:
         | According to the article, you can just have another LLM
         | summarise Gemini's watermarked output and that will "likely"
         | defeat the watermark detection.
        
           | scarmig wrote:
           | But, if all the good models can only be trained by large mega
           | corps with close connections to the government, it's only a
           | matter of time until that other LLM will just add its own
           | watermark.
        
         | onion2k wrote:
         | People use Google Search despite it being littered with adverts
         | and tracking. Maybe Google are counting on either being better
         | than the competition despite watermarking, or simply accepting
         | that people who don't care are enough of a market that it's
         | still worth adding.
        
         | beepbooptheory wrote:
         | Why does the user care if its watermarked? Surely there are
         | only some use cases for this stuff where it matters. Most of
         | the time isn't it just people having ephemeral chats where this
         | wouldn't matter?
        
           | ajdlinux wrote:
           | Using LLMs to write your essays and reports for school or
           | uni, in a way that could get you punished if caught, is a
           | reasonably big use case.
        
             | highcountess wrote:
             | I see no scenario where there won't be an LLM that is
             | deliberately tailored for that purpose, possibly even built
             | by an "intel" agency for the very purpose of having
             | blackmail over someone that may become useful later in
             | their career.
        
             | beepbooptheory wrote:
             | Agreed its probably a big use case in general, but like
             | token per token I bet its relatively small! How many big
             | papers do you have to write a semester? Even if its four,
             | that's nothing compared to the everyday use you will make
             | of it.
        
           | tokioyoyo wrote:
           | AIs and LLMs have an extremely uphill PR battle to fight
           | right now. Anything that is deemed AI generated is assumed to
           | be borderline trash (lots of exceptions, but you get the
           | point). So, I can see that if someone uses LLM to generate
           | text, they don't want it to be marked as "low effort
           | content".
        
             | beepbooptheory wrote:
             | There are definitely exceptions, and that there are maybe
             | proves that it is less Anti-AI prejudice at play and more
             | just reacting to things that are indeed trashy. It just so
             | happens a lot of it today is from AI I think (for, I hope,
             | obvious reasons).
             | 
             | Just to say, maybe give it a little time, but a watermark
             | like this is not going be thing that decides someone's
             | reaction in the near future, just what it says. (I am just
             | betting here).
             | 
             | But its going to be an uphill battle either way if you are
             | really getting the model to write everything, I do not envy
             | that kind of project.
        
         | nicce wrote:
         | Correct me if I'm wrong, but watermarking is only possible, if
         | the model has a limited set of input you can provide (affects
         | for the output) and a limited set of output it produces, and it
         | should be completely deterministic. And you should pre-
         | calculate all possible combinations.
         | 
         | And this should be also the case for every possible LLMs; then
         | you can compare which LLMs could produce which outputs based on
         | what inputs. Then there is some certainty that this output is
         | produced by this LLM and this another LLM might produce it as
         | well with these inputs.
         | 
         | So... impossible?
        
         | dartharva wrote:
         | If Google locks in enterprise clients using Google Workspace to
         | Gemini then they won't really have a choice. It is selling it
         | as an "add-on" already:
         | https://workspace.google.com/solutions/ai/#plan
         | 
         | Suffice to say it is evident that no other LLM will come close
         | in integration with Google Docs and other Workspace apps as
         | Gemini.
        
         | glenstein wrote:
         | People made this same argument about DRM escalations, about
         | increasing privacy violations in the browser, and about
         | Google's donations to support climate change misinformation.
         | Even about Facebook interface redesigns. Every variation of
         | "people will be driven to do X" I've ever heard assumes some
         | coherence and unity of collective purpose that rarely matches
         | the reality of how people behave.
         | 
         | There are counter examples, e.g. Unity. But catching that
         | lightning in a bottle is rare and merits special explanation
         | rather than being assumed.
        
           | tokioyoyo wrote:
           | Using LLMs in exams and homeworks has a different driver.
           | Getting caught results in punishment, so using alternative
           | would be better. None of the aforementioned examples have a
           | "stick" aspect to it when you stick to Google.
        
       | mateus1 wrote:
       | Google is branding this in a positive light but this is just AI
       | text DRM.
        
         | fastball wrote:
         | I for one am glad we might have a path forward to filtering out
         | LLM-generated sludge.
        
           | pyrale wrote:
           | > we
           | 
           | If by "we" you mean anyone else than Google and the select
           | few other LLM provider they choose to associate with, I'm
           | afraid you're going to be disappointed.
        
         | sebstefan wrote:
         | It's likely more about preventing model incest than digital
         | rights management
        
         | gwbas1c wrote:
         | Like all things a computer can / can't do; DRM isn't inherently
         | bad: It's how its used that's a problem.
         | 
         | IE, DRM can't change peoples' motivations. It's useful for
         | things like national security secrets and trade secrets, where
         | the people who have access to the information have very clear
         | motivations to protect that information, and very clear
         | consequences for violating the rules that DRM is in place to
         | protect.
         | 
         | In this case, the big question of if AI watermarking will work
         | / fail has more to do with peoples' motivations: Will the
         | general public accept AI watermarking because it fits our
         | motivations and the consequences we set up for AI masquerading
         | as a real person, or AI being used for misinformation? That's a
         | big question that I can't answer.
        
           | mateus1 wrote:
           | This is not a "good deed for the public" done by Google, this
           | is just a self serving tool to enforce their algorithms and
           | digital property. There is nothing "bad" here for the public
           | but it's certainly not good either.
        
       | FilipSivak wrote:
       | How is this supposed to work? By inserting special unicode
       | characters?
       | 
       | How can you watermark text?
        
         | hiatus wrote:
         | You can insert known spelling errors, choose certain phrasings,
         | and more. It doesn't have to be new characters added to the
         | text. Government security services have done stuff like this
         | for decades to weed out moles.
        
           | luigibosco wrote:
           | moles should know better than to utilize mountweazels!
           | https://en.wikipedia.org/wiki/Fictitious_entry
        
         | zorked wrote:
         | We've been studying unintentional watermarks for years.
         | 
         | https://en.wikipedia.org/wiki/Stylometry
        
         | das_keyboard wrote:
         | > SynthID-Text works by discreetly interfering in the
         | generation process: It alters some of the words that a chatbot
         | outputs to the user in a way that's invisible to humans but
         | clear to a SynthID detector. "Such modifications introduce a
         | statistical signature into the generated text," [...] "During
         | the watermark detection phase, the signature can be measured to
         | determine whether the text was indeed generated by the
         | watermarked LLM."
        
         | a2128 wrote:
         | I haven't read how Google is doing it, but one way it could be
         | done is to nudge which tokens get sampled. For example, every
         | other token could have an odd numbered id (where each token is
         | assigned an id from 0 to 32000 or however many it has). Then in
         | order to detect the watermark you just tokenize the text and
         | see if the pattern is there. A problem with this approach is
         | that it harms the accuracy and coherency, for example if you
         | ask "What is 2+2", and the token "4" is token #102, and it has
         | to pick an odd-numbered token, then it may respond with a wrong
         | answer or yap on strangely due to its limited selection of
         | tokens (like "The accurate answer to your mathematical query is
         | the number Four")
        
         | sebstefan wrote:
         | There's an article from ieee that explains it:
         | 
         | https://spectrum.ieee.org/watermark#:~:text=How%20Google%E2%...
        
         | sumtechguy wrote:
         | You do not even need extra characters (although they help). You
         | can use spaces, missing punctuation, upper/lower case in
         | particular cases, conjunction usage and not using it, word
         | substitution, common misspellings, transposed letters, etc. How
         | many extra spaces/tabs can you add to the end of a paragraph?
         | At the beginning? Between sentences? Inside them? Then you have
         | an AI agent design it and then train another one to detect it.
        
         | voidUpdate wrote:
         | As stated in the article, it alters the probabilities that the
         | network produces in a predictable way so that a different (but
         | still correct-sounding) word is picked. It subtly alters the
         | wording from what it would have output normally in such a way
         | that you can detect it, while still sounding correct to the
         | user
        
       | playingalong wrote:
       | > It has also open-sourced the tool and made it available to
       | developers and businesses, allowing them to use the tool to
       | determine whether text outputs have come from their own large
       | language models (LLMs), the AI systems that power chatbots.
       | However, only Google and those developers currently have access
       | to the detector that checks for the watermark.
       | 
       | These two sentences next to each other don't make much sense. Or
       | are misleading.
       | 
       | Yeah. I know. Only the client is open source and it calls home.
        
         | falcor84 wrote:
         | Is there significant throttling to prevent us from training a
         | classification model against it?
        
       | namanyayg wrote:
       | "An LLM generates text one token at a time. These tokens can
       | represent a single character, word or part of a phrase. To create
       | a sequence of coherent text, the model predicts the next most
       | likely token to generate. These predictions are based on the
       | preceding words and the probability scores assigned to each
       | potential token.
       | 
       | For example, with the phrase "My favorite tropical fruits are
       | __." The LLM might start completing the sentence with the tokens
       | "mango," "lychee," "papaya," or "durian," and each token is given
       | a probability score. When there's a range of different tokens to
       | choose from, SynthID can adjust the probability score of each
       | predicted token, in cases where it won't compromise the quality,
       | accuracy and creativity of the output.
       | 
       | This process is repeated throughout the generated text, so a
       | single sentence might contain ten or more adjusted probability
       | scores, and a page could contain hundreds. The final pattern of
       | scores for both the model's word choices combined with the
       | adjusted probability scores are considered the watermark. This
       | technique can be used for as few as three sentences. And as the
       | text increases in length, SynthID's robustness and accuracy
       | increases."
       | 
       | Better link: https://deepmind.google/technologies/synthid/
        
         | bgro wrote:
         | Couldn't this be easily disrupted as a watermark system by
         | simply changing the words to interfere with the relative
         | checksum?
         | 
         | I suspect sentence structure is also being used or, more
         | likely, the primary "watermark". Similar to how you can easily
         | identify if something is at least NOT a Yoda quote based on it
         | having incorrect structure. Combine that with other negative
         | patterns like the quote containing Harry Potter references
         | instead of Star Wars, and you can start to build up a profile
         | of trends like this statement.
         | 
         | By rewriting the sentence structure and altering usual wording
         | instead of directly copying the raw output, it seems like you
         | could defeat any current raw watermarking.
         | 
         | Though this hasn't stopped Google and others in the past using
         | bad science and stats to make unhinged entitled claims like
         | when they added captcha problems everybody said would be
         | "literally impossible" for bots to solve.
         | 
         | What a surprise how trivial they were to automate and the data
         | they produce can be sold for profit at the expense of mass
         | consumer time.
        
           | scarmig wrote:
           | In principle, it seems like you could have semantic
           | watermarking. For instance, suppose I want a short story.
           | There are lots of different narrative and semantic aspects of
           | it that each carry some number of bits of information:
           | setting, characters, events, and those lay on a probability
           | distribution like anything else. You just subtly shift the
           | probability distribution of those choices, and then it's
           | resistant to word choice, reordering, and any transformation
           | that maintains its semantic meaning.
        
         | baobabKoodaa wrote:
         | I'm fascinated that this approach works _at all_ , but that
         | said, I don't believe watermarking text will ever be practical.
         | Yes, you can do an academic study where you have exactly 1
         | version of an LLM in exactly 1 parameter configuration, and you
         | can have an algorithm that tweaks the logits of different
         | tokens in a way that produces a recognizable pattern. But you
         | should note that the pattern will be recognizable only when the
         | LLM version is locked and the parameter configuration is
         | locked. Which they won't be in the real world. You will have a
         | bunch of different models, and people will use them with a
         | bunch of different parameter combinations. If your "detector"
         | has to be able to recognize AI generated text from a variety of
         | models and a variety of parameter combinations, it's no longer
         | going to work. Even if you imagine someone bruteforcing all
         | these different combos, trouble is that some of the combos will
         | produce false positives just because you tested so many of
         | them. Want to get rid off those false positives? Go ahead, make
         | the pattern stronger. And now you're visibly altering the
         | generated text to an extent where that is a quality issue.
         | 
         | In summary, this will not work in practice. Ever.
        
       | rany_ wrote:
       | I really want to be able to try Gemini without the AI watermark.
       | IIRC they've used SynthID from the start and it makes me wonder
       | if it's the source of all of Gemini's issues.
       | 
       | Obviously Google claims that it doesn't cause any issues but I'd
       | think that OpenAI and other competitors would have something
       | similar to SynthID if it didn't impact performance.
        
       | js8 wrote:
       | I think people are already doing that. I frequently hear people
       | watermarking their speeches with phrases like "are we aligned on
       | this?", or "let's circle back" and similar.
        
         | lcnPylGDnU4H9OF wrote:
         | I can't tell if this is satire but that's just corp-speak. I
         | imagine those people also occasionally suggest "touching base"
         | and "taking this offline".
         | 
         | The phrases usually mean something useful, if one knows the
         | meaning, but it is amusing how much people seem to stick with
         | the same ones, even across companies.
        
           | js8 wrote:
           | I am not sure whether it was satire. I personally don't like
           | corp speak - it feels like people talking like that are not
           | humans. I am not sure I would welcome our AI overlords
           | speaking like this, either.
           | 
           | But I find the idea that people will subconsciously start
           | copying AI speech patterns (perhaps as a signal of
           | submission) amusing. I think it's gonna throw a wrench into
           | the idea.
           | 
           | IMHO LLMs either should help us communicate more clearly and
           | succinctly, or we can use them as tools for creativity
           | ("rephrase this in 18th century English"). Watermarking
           | speech sabotages both of these use cases.
        
       | playingalong wrote:
       | > the team tested it on 20 million prompts given to Gemini. Half
       | of those prompts were routed to the SynthID-Text system and got a
       | watermarked response, while the other half got the standard
       | Gemini response. Judging by the "thumbs up" and "thumbs down"
       | feedback from users, the watermarked responses were just as
       | satisfactory to users as the standard ones.
       | 
       | Three comments here:
       | 
       | 1. I wonder how many of the 20M prompts got a thumbs up or down.
       | I don't think people click that a lot. Unless the UI enforces it.
       | I haven't used Gemini, so I might be unaware.
       | 
       | 2. Judging a single response might be not enough to tell if
       | watermarking is acceptable or not. For instance, imagine the
       | watermarking is adding "However," to the start of each paragraph.
       | In a single GPT interaction you might not notice it. Once you get
       | 3 or 4 responses it might stand out.
       | 
       | 3. Since when Google is happy with measuring by self declared
       | satisfaction? Aren't they the kings of A/B testing and high
       | volume analysis of usage behavior?
        
         | varispeed wrote:
         | > I don't think people click that a lot.
         | 
         | I sometimes do, but I almost always give wrong answer or
         | opposite answer where possible.
        
           | froh wrote:
           | but why? what for?
        
             | thebruce87m wrote:
             | My timesheet SAAS constantly asks for feedback, which I
             | give 0/10 as constantly asking for feedback really annoys
             | me.
             | 
             | They then contact me and ask me why, so I tell them then
             | they say there is nothing they can do. A week later I'll
             | get a pop up asking for feedback and we go round the same
             | loop again.
        
             | varispeed wrote:
             | Because companies like Google are a cancer and I don't want
             | to give them data they didn't pay for.
        
               | froh wrote:
               | hm
               | 
               | reminds me of "what have the romans ever done for us?"
               | 
               | but thx for elaborating.
        
       | bko wrote:
       | This article goes into it a little bit, but an interview with
       | Scott Aaronson goes into some detail about how watermarking
       | works[0].
       | 
       | He's a theoretical computer scientist but he was recruited by
       | OpenAI to work on AI safety. He has a very practical view on the
       | matter and is focusing his efforts on leveraging the
       | probabilistic nature of LLMs to provide a digital undetectable
       | watermark. So it nudges certain words to be paired together
       | slightly more than random and you can mathematically derive with
       | some level of certainty whether an output or even a section of an
       | output was generated by the LLM. It's really clever and
       | apparently he has a working prototype in development.
       | 
       | Some work arounds he hasn't figured out yet is asking for an
       | output in language X and then translating it into language Y. But
       | those may still be eventually figured out.
       | 
       | I think watermarking would be a big step forward to practical AI
       | safety and ideally this method would be adopted by all major
       | LLMs.
       | 
       | That part starts around 1 hour 25 min in.
       | 
       | > Scott Aaronson: Exactly. In fact, we have a pseudorandom
       | function that maps the N-gram to, let's say, a real number from
       | zero to one. Let's say we call that real number ri for each
       | possible choice i of the next token. And then let's say that GPT
       | has told us that the ith token should be chosen with probability
       | pi.
       | 
       | https://axrp.net/episode/2023/04/11/episode-20-reform-ai-ali...
        
         | littlestymaar wrote:
         | Sounds interesting, but it also sounds like something that
         | could very well be circumvented by using a technique similar to
         | speculative decoding: you use the censored model like you'd use
         | the fast llm in speculative decoding, and you check whether the
         | other model agrees with it or not. But instead of correcting
         | the token every time both models disagree like you'd do with
         | speculative decoding, you just need to change it often enough
         | to mess with the watermark detection function (maybe you'd
         | change every other mismatched token, or maybe one every 5
         | tokens would be enough to reduce the signal-to-noise ratio
         | below the detection threshold).
         | 
         | You wouldn't even need to have access to an unwatermarked
         | model, the "correcting model" could even be watermaked itself
         | as long as it's not the same watermarking function applied to
         | both.
         | 
         | Or am I misunderstanding something?
        
         | nicce wrote:
         | I don't think that provable watermarking is possible in
         | practice. The method you mention is clever, but before it can
         | work, you would need to know the probability of the every other
         | source which could also be used to generate the output for the
         | same purpose. If you can claim that the probability of that
         | model is much higher on that model than in any other place,
         | including humans, then watermark might give some stronger
         | indications.
         | 
         | You would also need to define probability graph based on the
         | output length. The longer the output, more certain you can be.
         | What is the smallest amount of tokens that cannot be proved at
         | all?
         | 
         | You would also need include humans. Can you define that for
         | human? All LLMs should use the same system uniformally.
         | 
         | Otherwise, "watermaking" is doomed to be misused and not being
         | reliable enough. False accusations will be take a place.
        
       | tiffanyh wrote:
       | OT: The publication (Spectrum by IEEE) has some really good
       | content.
       | 
       | It's starting to become a common destination for when I want to
       | read about interesting things.
        
       | samatman wrote:
       | This is information-theoretically guaranteed to make LLM output
       | worse.
       | 
       | My reasoning is simple: the only way to watermark text is to
       | inject some relatively low-entropy signal into it, which can be
       | detected later. This has to a) work for "all" output for some
       | values of all, and b) have a low false positive rate on the
       | detection side. The amount of signal involved cannot be subtle,
       | for this reason.
       | 
       | That signal has a subtractive effect on the predictive-output
       | signal. The entropy of the output is fixed by the entropy of
       | natural language, so this is a zero-sum game: the watermark
       | signal will remove fidelity from the predictive output.
       | 
       | This is impossible to avoid or fix.
        
       | espadrine wrote:
       | The academic paper:
       | https://www.nature.com/articles/s41586-024-08025-4
       | 
       | They use the last N prefix tokens, hash them (with a keyed hash),
       | and use the random value to sample the next token by doing an
       | 8-wise tournament, by assigning random bits to each of the top 8
       | preferred tokens, making pairwise comparisons, and keeping the
       | token with a larger bit. (Yes, it seems complicated, but
       | apparently it increases the watermarking accuracy compared to a
       | straightforward nucleus9 sampling.)
       | 
       | The negative of this approach is that you need to rerun the LLM,
       | so you must keep all versions of all LLMs that you trained,
       | forever.
        
       ___________________________________________________________________
       (page generated 2024-11-05 23:00 UTC)