[HN Gopher] Not all tokens are meant to be forgotten
       ___________________________________________________________________
        
       Not all tokens are meant to be forgotten
        
       Author : MarcoDewey
       Score  : 48 points
       Date   : 2025-06-04 23:15 UTC (23 hours ago)
        
 (HTM) web link (arxiv.org)
 (TXT) w3m dump (arxiv.org)
        
       | pixl97 wrote:
       | > However, they tend to memorize unwanted information, such as
       | private or copyrighted content,
       | 
       | I mean humans don't forget copyrighted information. We just
       | typically adjust it enough (some of the time) to avoid getting a
       | copyright strike while modifying it in some way useful.
       | 
       | We don't forget 'private' information either. We might not tell
       | other people that information, but it still influences our
       | thoughts.
       | 
       | The idea of a world where we have AI minds forget vast amounts of
       | information that humans have to deal with every day is concerning
       | and dystopian to me.
        
         | johnjreiser wrote:
         | I'd counter with an anecdote; I had a colleague that boasted
         | how he memorized a classmate's SSN in college and would greet
         | him by SSN when seeing him years later. Is the goal of AI to
         | replicate the entirety of the human experience (including
         | social pressures, norms, and shame) or a tool to complement
         | human decision making?
         | 
         | While, yes, you can argue the slippery slope, it may be
         | advantageous to flag certain training material as exempt. We as
         | humans often make decisions without perfect knowledge, and
         | "knowing more" isn't a guarantee that it produces better
         | outcomes, given the types of information consumed.
        
           | lmm wrote:
           | Knowing more might not improve your accuracy but it's not
           | going to harm it. Forcibly forgetting true parts of your
           | knowledge seems far more likely to have unintended
           | consequences.
        
             | conception wrote:
             | Counterpoint: There are plenty examples of breakthroughs
             | from folks who are ignorant of the "right" way to go about
             | it. A fresh take isn't always bad.
        
             | Dylan16807 wrote:
             | I disagree. Actively fighting against your memory will slow
             | you down in any context where some memorized idea is
             | similar to what you're doing but you shouldn't be using the
             | memorized idea.
        
             | lou1306 wrote:
             | One obvious consequence: the model might still produce
             | copyright infringement because it thinks its creative ideas
             | are novel.
        
               | genewitch wrote:
               | If the copyrighted content is not in the training data,
               | and I mean explicitly, and the AI produces a copyrighted
               | output, I'd argue it's a clean room re-implementation,
               | and also it ought devalue the original work, moreso if
               | the work is more recent. Maybe.
               | 
               | I get that "first to publish" matters to a lot of people,
               | but, say 5 unrelated people are writing unique
               | screenplays about a series of events that seems important
               | to them or culture or whatever; if they all come up with
               | very similar plots and locations and scenes, it just
               | means that the idea is more obvious than non-obvious.
               | 
               | Please, argue. I haven't fully reconciled a lot of this
               | to myself, but off the cuff this'll do.
               | 
               | The logic being - if an AI without taint produces some
               | other work, that work drew on the same information the
               | model did, and came to the same "conclusion" - which
               | means with a time machine, you could wipe the LLM, go
               | back to the period of the original work, train the LLM,
               | and produce the work contemporaneous to the original.
               | Hope that made sense.
        
               | lmm wrote:
               | > If the copyrighted content is not in the training data,
               | and I mean explicitly, and the AI produces a copyrighted
               | output, I'd argue it's a clean room re-implementation
               | 
               | You can't claim it's a clean room without actually doing
               | the legwork of making a clean room. Not including the
               | copyrighted work verbatim isn't enough, you would need to
               | show that the AI hadn't seen anything derived from that
               | copyrighted work, or that it had seen only non-
               | copyrightable pieces.
        
               | lou1306 wrote:
               | > The logic being - if an AI without taint produces some
               | other work, that work drew on the same information the
               | model did, and came to the same "conclusion" - which
               | means with a time machine, you could wipe the LLM, go
               | back to the period of the original work, train the LLM,
               | and produce the work contemporaneous to the original.
               | Hope that made sense.
               | 
               | This logic would immediately get shot down by an
               | "Objection, speculation" in an actual litigation.
               | Besides, the technicalities of how the work was produced
               | don't really play a role in assessing infringement. PK
               | Dick wrote "The man in the high castle" by extensively
               | using the I Ching, but if I use it and recreate the novel
               | by complete accident I would still be infringing.
               | 
               | By the way, I highly suggest Borges's "Pierre Menard,
               | Author of the Quixote" as a great story on the topic of
               | authorship :)
        
               | genewitch wrote:
               | > PK Dick wrote "The man in the high castle" by
               | extensively using the I Ching, but if I use it and
               | recreate the novel by complete accident I would still be
               | infringing.
               | 
               | I touched on this, with the comment that we love "first
               | to market." That multiple people coming up with the same
               | output may mean that the idea isn't that novel. whether
               | that matters or not isn't really relevant to me.
               | 
               | The part you quoted was just a thought experiment to
               | explain why i compared it to a "clean room
               | implementation" - note it also avoids this argument from
               | a sibling comment:
               | 
               | >need to show that the AI hadn't seen anything derived
               | from that copyrighted work
               | 
               | since there could not possibly be any derived work prior
               | to the "original" work being published. For the sake of
               | argument.
        
               | kgwgk wrote:
               | I highly suggest Borges's "Pierre Menard, Author of the
               | Quixote" as a great story on the topic of authorship :)
        
               | JadeNB wrote:
               | The repetition of the end of lou1306's comment
               | (https://news.ycombinator.com/item?id=44190054) "By the
               | way, I highly suggest Borges's 'Pierre Menard, Author of
               | the Quixote' as a great story on the topic of authorship
               | :)" has to be a joke ... right?
        
               | kgwgk wrote:
               | Good question! Is Pierre Menard's Quixote a repetition of
               | Cervantes' or is it a completely different work that just
               | happens to contain the same words?
        
               | JadeNB wrote:
               | > Is Pierre Menard's Quixote a repetition of Cervantes'
               | or is it a completely different work that just happens to
               | contain the same words?
               | 
               | I think that that is not the right question. It is a
               | repetition of Cervantes's work by design, at least if one
               | takes, as I do, 'repetition' to mean saying or writing
               | the same words in the same order. I think the question is
               | whether it is therefore the same work, or a different
               | work that contains the same words.
        
           | lynx97 wrote:
           | The goal of AI is to make money. All the moralisation is very
           | human, but also extremely naive.
           | 
           | BTW, I don't really understand what "social pressure" and
           | "shame" has to do with your story? In my book, the person
           | with a good memory isn't to blame. They're just demonstrating
           | a security issue, which is a good thing.
        
             | falcor84 wrote:
             | In that example, the mnemonist should be demonstrating the
             | security issue to the government, and not to their friend.
             | We have social taboos for this reason. As an extreme
             | example, I wouldn't greet a person by their penis size
             | after noticing it in the locker room - some information
             | should still be considered private, regardless of how we
             | came to obtain it.
             | 
             | Same with an LLM, when it got sensitive information in its
             | weights, regardless of how it obtained it, I think we
             | should apply pressure/shame/deletion/censorship (whatever
             | you call it) to stop it from using that information in any
             | future interactions.
        
               | lynx97 wrote:
               | I am probably too autistic to recognize remembering a
               | personal datum as a taboo.
               | 
               | However, I am totally on your side regarding LLMs
               | learning data they shouldn't have seen in the first
               | place. IMO, we as a society are too much chicken to act
               | on the current situation. Its plain insane that everyone
               | and their dog knows that libgen has been used to train
               | models, and the companies who did this experiencing NO
               | consequences at all. After that, we shouldn't be surpised
               | if things go downhill from here on.
        
         | squidbeak wrote:
         | I agree. As far as copyrighted and artistic works go, I've
         | never fully understood what the objection is. If the work is
         | being remixed not copied then it surely falls under fair use?
         | Meanwhile, if it creates something new in an artist's style,
         | it's only doing what talented imitators routinely do. There's
         | the economic argument. But if that's accepted, then for
         | fairness it would have to be extended to every other profession
         | which stands to be wiped out by AI, which would be daft.
         | 
         | New works in familiar styles are something I can't wait for.
         | The idea that the best Beethoven symphony hasn't been composed
         | yet, or that the best Basquiat hasn't been painted yet, or that
         | if the tech ever gets far enough, Game of Thrones might
         | actually be done properly with the same actors, is a pretty
         | mouthwatering prospect. Also styles we haven't discovered, that
         | AI can anticipate. How's it to do that without a full
         | understanding of culture? Hobbling the delight it could bring
         | generally for the sake of protected classes will just make the
         | tech less human and a lot less exciting.
        
           | wizardforhire wrote:
           | Mind if I ask a few questions? Whats your current address,
           | dob, ssn or NINO or equivalent, your full legal name, mothers
           | maiden name, fathers place of birth, mothers place of birth,
           | country of origin, do you drive? Whats your license number?
           | How about a bank? Could I have your account and routing
           | number as well as the answers to any security questions? How
           | about investments I'm gonna need your accounts and passwords
           | for these as well...
           | 
           | > As far as copyrighted and artistic works go, I've never
           | fully understood what the objection is ... > But if that's
           | accepted, then for fairness it would have to be extended to
           | every other profession which stands to be wiped out by AI,
           | which would be daft. ... > Hobbling the delight it could
           | bring generally for the sake of protected classes will just
           | make the tech less human and a lot less exciting.
           | 
           | So let me get this straight, you want to ruin the livelihoods
           | of everyone so you can have a fancier toy to play with?
           | 
           | When your life is ruined and can't make a living you'll have
           | the answers you desire and understand the objections to why
           | you can't have fancier toys.
           | 
           | But heres the thing, and with the way the world is going atm,
           | not being able to make a living is going to be the least of
           | your and everyone else's worries that feel the way you do if
           | ya'll get your way.
           | 
           | People don't like having their livelihoods taken away, and
           | when you threaten the livelihoods of their children... people
           | tend towards violence.
           | 
           | I really wish there was a more polite way to put this. Alas
           | what you're proposing is all out war for what? A better game
           | of thrones?
        
             | squidbeak wrote:
             | Violent artists with pitchforks, eh? Aside from their
             | supposed predisposition to vengeful bloodlust, is there any
             | other reason these protected classes should enjoy a
             | different status to any other worker?
        
           | wat10000 wrote:
           | If it's remixed then it would be a derivative work and you'd
           | need permission from the original copyright holder, just like
           | if you literally remixed a song, or made a movie based on a
           | novel.
           | 
           | IMO the only reason there's even a question about whether
           | LLMs can legally be trained on copyrighted works without
           | permission is that the training is being done by (agents
           | working on behalf of) rich people. If you or I scraped up
           | every copyrighted work we could get our hands on without ever
           | asking permission, trained an LLM on it, and then tried to
           | sell access to the result? Just ask Aaron Swartz how that
           | sort of thing goes, and his actions were orders of magnitude
           | less.
           | 
           | Humans don't forget copyrighted material but we also don't
           | normally memorize it. It takes substantial time and effort to
           | be able to reproduce copyrighted material with just your
           | brain.
        
       | JimDabell wrote:
       | There's a related paper that Meta published a couple of days ago
       | that is worth looking at:
       | 
       | > How much do language models memorize?
       | 
       | -- https://arxiv.org/abs/2505.24832
       | 
       | -- https://news.ycombinator.com/item?id=44171363
       | 
       | It shows that models are limited in how much they can memorise
       | (~3.6 bits per parameter), and once that threshold is reached,
       | the model starts to generalise instead of memorise.
        
       ___________________________________________________________________
       (page generated 2025-06-05 23:02 UTC)