I'm sure we all have our own feelings about IP law, but remember what happens to regular people who try stuff like this. I don't think the RIAA, Disney, or Nintendo (or the government) are going to be pleased to hear "it's not piracy! It's a transformative experience protected by fair use!"
> Humans are allowed to "absorb" art around them into their brains and generate derivative art. People may copy Miyazaki's style... why shouldn't an AI farm be allowed to?
Let's put aside for a moment that AI may have "consumed" some art without a license (e.g., "google books" - did google purchase every book?).
if anything, I'd campaign for "we should limit copyright because it already doesn't work for Ai"
Copying Miyazaki's style ... or copying Monet's style... those aren't fan art.
People may take a penny from the tray at the 7-11, so why can't an AI farm take pennies from all the trays? Or take them from a much bigger tray and do it a couple of million times?
No one is stopping you from using similar proportions or colors as Miyazaki to draw a character. You are also allowed to draw your own interpretation of an electric mouse-like monster.
Copyright infringement occurs if that character looks exactly like say Totoro or Pikachu. That is not “in the style of”, that is copying.
A problem with LLMs is that since their corpus is so large, it is difficult to identify when any given output is crossing that line because a single observer’s knowledge of the works influencing the output is limited. You might feed it a picture of your grandfather and it returns an almost exact copy of a grandfather character from a Miyazaki film you haven’t seen. If you don’t share the output with others, it might never be noticed that the infringement occurred.
The given argument conflates the slightest influence with direct copying. It is a reductive take that, personally, I’ve found emblematic of pro-LLM arguments.
I don't like the idea that photos I've published on, say, flickr have been pulled into these. Especially stuff I've published with creative commons non-commercial use.
That narrative only got picked up because people needed a reason to demonize evil corps that they already hated for unrelated reasons.
[1] Yes, if you create "new" works from your learning that are basically copies, that has always been infringement. I'm talking about the general case.
Either they aren't evil in which case they're being demonized, or they're already evil in which case demonization is redundant.
Keeping aside the motives of people, what is clear is that scale effects of AI cannot be ignored. An AI "learning" millions of pieces of content in a short span is not the same as humans spending time, effort and energy to replicate someone's style. You can argue that its 'neural nets' in both cases, but the massive scale is what separates the two.
A village is not a large family, a city is not a large village, ... and all that.
If you were trying to be charitable rather than clever, you would have read "evil corps" as "corps that the critic regards as evil".
>Keeping aside the motives of people, what is clear is that scale effects of AI cannot be ignored.
Okay, so just give some kind of standard -- any clear, articulable standard -- for how and why the scale matters. It's a cop-out to just rest your case on a hand-wavy "it changes at scale".
The term "learning" (I presume from "machine learning") shoulders a lot of weight. If we describe the situation more precisely, it involves commercially exploiting literature and other text media to produce a statistical corpus of texts, which is then commercially exploited. It's okay if that is licensed, but none of the AI companies bothered to license said original texts. Some (allegedly) just downloaded torrents of books, which is clear as day piracy. It has little to do with "learning" as used in common English — a person naturally retaining some knowledge of what they've consumed. Plain English "learning" doesn't describe the whole of what's happening with LLMs at all! It's a borrowed term, so let's not pretend it isn't.
What's happening is closer to buying some music cassettes, ripping parts of songs off them into various mixtapes, and selling them. The fact that the new cassettes "learned" the contents of the old ones, or that the songs are now jumbled up, doesn't change that the mixtape maker never had a license to copy the bits of music for commercial exploitation in the first place. After the infringement is done, the rest is smoke and mirrors...
It's "commercially exploiting literature" in the same sense that an author would if they read a bunch of novels and then wrote their own based on what the learned from the pre-existing text. The whole point in dispute is whether that turns into infringement when an AI does it.
By labeling only one of them as "commercially exploiting literature" but not the other, you're failing to distinguish them in any meaningful way, and basically arguing from name-calling.
>It has little to do with "learning" as used in common English — a person naturally retaining some knowledge of what they've consumed. Plain English "learning" doesn't describe the whole of what's happening with LLMs at all! It's a borrowed term, so let's not pretend it isn't.
That's fair, that you can't just call them both "learning" and call it a day. But then the burden's on you to show how machine learning breaks from the time-honored tradition of license-free learning/"updating what you write based on having viewed other works". What's different? What is it about machine learning that makes it infringement in a way that it isn't when humans update their weights from having seen copyrighted works?
>What's happening is closer to buying some music cassettes, ripping parts of songs off them into various mixtapes, and selling them. The fact that the new cassettes "learned" the contents of the old ones, or that the songs are now jumbled up, doesn't change that the mixtape maker never had a license to copy the bits of music for commercial exploitation in the first place.
Okay, but (as above) to make that case, you'd need to identify where "acceptable" learning/"updating what you write based on having viewed other works" crosses over into the infringing mixtape example, and I have yet to see anyone try beyond "they're evil corps, it must be bad somehow".
Human learning doesn't involve making a copy (or any other use of an exclusive rights) as defined in copyright law (the human brain not being a fixed medium), AI training does, because digital storage is.
AI training may fall into the Fair Use exception in the US, but it absolutely does not fall through the same gap that makes human learning not even eequire fair use analysis since it doesn't meet the definitions ser out for a violation in the first place.
That's just false -- AI models themselves only store parameter weights, which represent a high-level, aggregated understanding across all data that was learned on, i.e. what human brains do. This is clear from all the examples where you have to painstakingly trick them into producing exact text.
And even if they did store something that's "effectively a copy", that's no more copyright infringement than when Google caches a site in order to handle search queries. It's not copyright infringement until they start redistributing [non-fair-use] content from the sites.
I mean sure, if you go to painstaking lengths, you can trick your computer into making some noise that seems vaguely similar to the copyrighted work it was trained on, but I trust the consumer to make their own fair use evaluation.
Human learning doesn't involve fixing anything, whether or not it would be a copy, into a fixed medium, and therefore even the question of whether or not it would be a copy if it were in a fixed medium, much less the question of whether or not such a copy would be exempted due to fair use, does not arise.
AI training does involve fixing something into a fixed medium, which makes it disanalogous to human learning. This raises the question of whether the thing fixes (parameter weights) is or is not legally a lossy mechanical copy of the training corpus, which further potentially raises the question of whether the incorporation of individual copyright protected works into the training corpus combined with its use in training would (before considering exceptions like Fair Use) violate the copyright on the works involved and whether, if it would, the use nevertheless satisfies the requirements of one or more of those exceptions (Fair Use being the one usually argued for.)
> And even if they did store something that's "effectively a copy", that's no more copyright infringement than when Google caches a site in order to handle search queries.
Among the reasons that Google removed outside access to its cached copies is that the evolution of the web meant that providing them increasingly potentially had negative revenue on the original content providers, which harms the Fair Use case for caching, since effect on the market of the original work is a fair use factor. (Google's cache was ruled as fair use based on the factual situation -- including implied consent, which AI model trainers have a much weaker arguent for -- in 2006, but changes in the facts relevant to fair use can change the outcome.)
But AI training is not so analogous to Google's cache (much less the situation of Google's cache in 2006) that one can simply leap with no analysis from one being fair use to the other in the first place. That's applying wishful thinking, not Fair Use analysis.
No, it breaks one part of the analogy, which has never been considered relevant: how long the learning persists. Yes, computers can store weights of a AI model much longer than any human could. But storing this "impact of having viewed a copyrighted work" has never been a factor in considering something infringing, regardless of how long it's stored. Courts don't consider it infringement if you simply use what you have learned from reading previous novels (the updates to your brain's neural weights) in producing new content.
Your argument is saying that the "fixedness of storing model weights" (aggregated high level understanding) can cause it to be infringement. That's without precedent. It would imply that if you're really good at "fixing" (remembering) the style of an author you read 50 years ago, it somehow crosses over into infringement because you "fixed that understanding into a medium" (your brain). That's not how it works at all.
>Among the reasons that Google removed outside access to its cached copies
I wasn't referring to the cached part of a site that Google serves to users, but the undistributed cache that they hold merely to know which sites to point you to, so you're not addressing the analogy. Here Google does store an exact copy (of at least some portions) and even then it's not considered copyright infringement until they start redistributing that content (or at least, too much of it).
My point was that acceptance of this practice even further bolsters the case that AI models aren't infringing, because, even if they did store exact copies, that generally not considered infringement until they start serving close-enough copies of the original copyrighted content.
And, while copyright prohibits some sorts of reproduction of copyrighted materials, it doesn't give rightsholders veto power over all downstream uses of legal copies.
Let's just flag anything that gets in the way of profit and peace of mind.
Truth will find you.
minimaxir•6mo ago
JonChesterfield•6mo ago
Also Google wouldn't give me an antonym for "satire", only the output of a LLM which thinks synonym is the same thing as antonym.
leephillips•6mo ago
Edit: I should have said not all nouns have antonyms.
monster_truck•6mo ago
Centigonal•6mo ago
gotoeleven•6mo ago