There may be a misalignment in intent of the claim and interpretation of the claim. As someone that researches generative modeling I actually think there is an important aspect to this question, but I do not think that this question has anything to do with how the brain or the machine learn art. It has to do with legality and morals.
So I'll break it down. We believe that it is morally and legally acceptable for a human to look at copyrighted artwork and even mimic it in the process of learning how to become a better artist (sales are where the morality breaks down and especially with impersonation). The question is "where is the nuanced difference between a machine using that data and a human using that data to learn?" This doesn't depend on the learning techniques just like how no one cares if one person learns differently than another person. Obviously no one thinks AI art should impersonate real artists nor do they think people should sell this work if it contains copyrighted material. That's in line with the human artist values (fine to draw Mickey Mouse, not fine to sell a drawing of Mickey Mouse and worse to sell that drawing and claim it is official Disney art).
This is a very important question because we need to create laws about how we can train these systems and how we handle the data that they produce (two very different things!). The line between human and machine is a lot thinner than people think (think digital painting and CGI), and it doesn't matter that stochastic algorithms learn differently than humans learn. The question is about how/if learning material can be used and if machines should be treated differently. And if so, why.
But this is way more than one sentence.
This is a rather prejudiced statement. AI does pose some interesting questions about copyright, but the most obvious question is whether the existing laws are sufficient. You go on to ask that question, but this statement presupposes the answer is that they’re not.
I assert that existing legal systems do not suffice. Human governments and courts have shown themselves to be near-sighted, slow, and even corrupt when it comes to rapidly advancing lucrative technology (except perhaps when that tech is a weapon).
If I say someone "voted in a prejudiced way," it's technically true if they'd already decided how they'd vote before they did so, but most people would understand my meaning to be that their vote was motivated by bigotry of some kind.
Oh, and I assert that anyone who doesn’t think our legal system can properly decide who if anyone is being wronged by these tools is just plain ignorant of the law.
Uber/AirBnB facilitating illegal activity but growing large enough to force legalization.
Facebook et al providing data and conduits to influence elections.Well here's some basics about the law. Broadly speaking, a trial requires a plaintiff who makes a claim of a wrong allegedly committed by a defendant.
In the United States, which at the federal and state level (spare Louisiana) follows English common law jurisprudence, there are two kinds of wrongs: Private and public. Broadly speaking, public wrongs are committed against "the community", while private wrongs are committed against an individual.
So if we're talking about say, Stable Diffusion, we can reference the actual claim of the private wrong being made and the details of the case. The plaintiffs are some artists. The defendants are some companies who created a ML model. The claim is that they were wronged for copyright infringement, DMCA violations, common law and statuary rights to publicity, etc. For copyright infringement, the defense will attempt to show that SD is fair use and their reference Sony v Universal, Google v Authors Guild, Baker v Selden, ABC v Aerokiller. The plaintiffs will argue it's not fair use while also attempting to show the damages to the marketplace for competing works. There will be strategies on both sides for the other claimed wrongs. I don't see anything that the legal system can't place mainly into existing frameworks while also adding to the existing corpus of case law and helping to further define what is and isn't considered a wrong for future courts when it comes to training ML models on published works.
So I'm not exactly sure what you're trying to say about Uber, AirBnB or Facebook because those just sound like a lot of political opinions as they are not legally framed. If you and other people want representatives to write new laws about social media companies, ride sharing services or short-term rentals, well go right ahead!
IMO, no new laws are necessary for the two "high-profile" (hah!) cases against Stable Diffusion and Copilot. I'm also pretty sure that the defendants will successfully prove to the courts that these tools are indeed fair use.
Politically, I would want my representatives to argue against legislation that would change such an interpretation!
Software is often built to perform well for expected inputs, yet when given unexpected inputs it may produce bizarre results that when passed to other software create difficult-to-anticipate cascades of failures. The legal system is capable of the same.
These are in fact quite similar questions. There is a nuanced aspect to this though: with ML generation you're less likely to be aware that you have produced copyrighted (or derivative) work or are plagiarizing. The reason the question is being asked is not to presuppose that the answer is that they're not but it is instead a response to those that are saying existing laws aren't enough.
It's trivial because the real question is "Can AI produce creative work which is at least culturally equivalent - and perhaps even better - than the best human creative work?"
I'm fairly sure the answer to that question is "Yes". Because as soon as you go past modelling content and start modelling psychology, aesthetics, politics, and the dynamics of culture itself, the question answers itself.
With fast-enough processing and a large enough dataset, why wouldn't Behavioural/Political/Cultural AI be possible, perhaps with different AI systems using culture and human affect itself as a memetic battlefield?
Of course then we're very much not in Kansas any more, Toto. The danger - actually the likelihood - is that AI becomes an irresistibly seductive mechanised sociopath, able to automate everything, including politics, culture, and everything else that makes humans human.
But what about ethics, you ask?
How much AI research is asking that question at all, never mind taking it seriously?
> How much AI research is asking that question at all, never mind taking it seriously?
I think quite a lot of us (speaking as a researcher) are in fact asking ethic questions. But I think there is nuance we're making decisions on that others don't get. I think there are also nuances that many artists get that we researchers don't. Unfortunately I think we're frequently speaking past one another, and worse, we like sensationalism and so this is even being encouraged. That makes us to talk in bubbles because we feel like we're not being heard and get understandably frustrated. But this is a societal issue and as such we need a lot of points of views to resolve and the bubbling effect will prevent us doing this in a useful manner. Communication in the global era is harder than we're primed to think because communication in local groups has far fewer issues and when there are issues they are often quickly resolved.
I agree, even computers can do that, but what is the license of the output? (Both human generated output and AI generated output.)
The problem is that someone made one of the Ai draw a few Mickey Mouse versions, and claimed that the license of the AI said that the output was public domain, so everyone can use the AI-created drawings of Mickey Mouse freely. (I guess Disney disagree.)
Expand this a little bit in the opposite direction…imagine if Stable Diffusion was made illegal. Someone accuses me of using this illegal tool for an image that doesn’t look like anyone else’s image as far as the court is concerned for copyright. I put the image on my website. If the image itself is not at all infringing, then what is the evidence that Stable Diffusion was used? Should the police be issued a warrant to search my private property for proof that I used Stable Diffusion without a shred of evidence?
This frames the conversation about copyright within the legal structure and not some unrelated philosophical structure.
I fully agree.
Anyway, I found the original link https://twitter.com/eze3d/status/1601695610498781184
It’s the kind of conversation to have in the context of art (better yet if your contribution to the conversation is art itself) or the philosophy of language, although the question could be a little less “stoned in a Freshman dorm” and more informed by the actual discourse of those specialties as these questions are sort of old hat.
If you want to ask meaningful questions of these tools and copyright you need to study the law, plain and simple.
That is, in a non ideal scenario, if it creates are that wouldn't normally be subject to copyright, then the process (call it learning, call it mapping, call it baking for all I care) is largely irrelevant.
In an ideal scenario, no artwork would be owned and we wouldn't have this discussion at all but alas.
I don't see how valuing AI output is some kind of nihilistic devaluing of the human race, unless you already hold some kind of preconceived negative idea of machine learning models.
Remember these models are built by humans. Its human ingenuity that's scaring the pants off the art industry at the moment. TALL men and women and others of INDUSTRY and BUSINESS are making COOL THINGS with VALUE.
It can go the other way too, like proof that diffusion models memorize heavily repeated training examples being used as evidence they aren't creating stuff at all and just copy pasting, but artists can memorize whole works too, especially artists that copy style really well.
Many English classes will have you verbatim memorize poetry as well, to show how things were done with memorization in oral storytelling traditions.
This proof actually demonstrated the opposite. In that paper it was found that something like only 11 pictures from millions were recoverable. They were only recoverable because those handful of pictures were accidentally duplicated in the training data many multiples more than other pictures. Even pictures which were accidentally duplicated dozens of times were found to be unrecoverable by the trained model.
It's an important observation that humans are just as capable of doing tasks without understanding them, and so its no surprise that the computer doesn't understand them either.
Not only is it less precise term, but it gives the wrong implications.
Personally, I'm on the side of releasing training data. Let everybody train on everything. But it's always felt absurd to say that the ML models are "learning" things.
But hey, none of us know how learning works anyway, right? So maybe it's not such a big distinction. As you say, none of us can pinpoint why a model isn't learning vs why we are.
It seems almost definitionally obvious that what an LLM does is not the same as what a human does – both on the basis that if all human writing were merely done via blending together other writing we had seen in the past, it would appear to be impossible for us to have developed written communication in the first place, and on the basis that when I write something, I mean something I am then attempting to communicate. An LLM never means to communicate anything, there is no there there; it simply reproduces the most likely tokens in response to a prompt.
To insist that we're just a bunch of walking, breathing prompt-reproducers essentially seems like it's rooted in a belief that we have no interior lives, and that meaning in writing or art is utterly illusory.
It’s not said very much, but this style of dehumanization is really corrosive in a way that directly benefits the worst forms of human governments and structures, and this fact goes i think genuinely unrecognized too often in tech-land.
if we really are p-zombies, then those people aren’t really suffering, right, so it’s fine …
Let’s assume humans are not just evolved pattern machines for a second. A human can still do a completely non profound work of art following a prompt to draw X in the style of Y. And that’s ok. So why can a machine not do the same?
Surely not everything a human does is intrinsically profound.
The original discussion was precisely an objection to the attitude underlying "How is *GPT taking in data and producing an output different than a human learning a skill and making prose/code/art?" and the answer is right in your premise - not everything a human does is not profound. A human can intend to mean something with prose or art, even if not all prose or art means something — but any meaning we see in ChatGPT’s output is essentially pareidolia.
However profundity doesn’t need to factor into the debate of whether ai should or should not be allowed to train on things. If we allow humans to copy things, then Humans ought to be allowed to copy things with dumb non sentient ai too.
Ai in the current state is just a tool, much like a paint brush.
Cue the inevitable appeal to copying exact works, rebuttals to training on human painted mimicries and then bam, you’ve got the authors special style learned by the model with extra steps.
It’s annoying and pointless.
Art that is merely visually intriguing is not very interesting. If an artist makes something without a particular idea to communicate, it’s just aesthetics. It is not profound. If an artist has an idea and creates a work that represents it, then maybe it is profound. But it doesn’t matter if it was made with paint or a computer. The idea is the profound thing. AI is not sentient. It’s still the user.
The appeals to pareidolia are wrong. Synthesis of ideas from past data is natural. But the AI does not choose things. What you’re really complaining about is creation of art from apparent randomness. Not the AI model alone but monkeys on a typewriter getting something compelling from the AI.
What do we do when the tools are so powerful that a monkey creates a profound work that the monkey doesn’t understand? Shrug.
> The appeals to pareidolia are wrong. Synthesis of ideas from past data is natural. But the AI does not choose things. What you’re really complaining about is creation of art from apparent randomness. Not the AI model alone but monkeys on a typewriter getting something compelling from the AI.
No, you've failed to understand what I'm saying entirely (because, again, you've responded to some other post that only exists in your mind).
What I'm talking about is intention and its relationship to meaning, in the philosophical sense (and not... copyright or whatever it is you're rambling on about).
Witness: when ChatGPT famously mis-asserts the number of characters in a word (say, that there are twelve characters in the word "thirteen"), it's not that it's trying and failing to count, because it's confused by letter forms or its attention wanders like a 3 year old or its internal representation of countable sets glitches around the number 8 or something – it never counted anything at all, it's simply the case that twelve is the most statistically likely set of tokens corresponding to that input prompt per its training set. And when it produces a factually correct result (say, "there are 81 words in the first sentence of the declaration of independence"), it produces it for exactly the same reason – not because it has counted the words and formed an internal representation and intends to mean its internal understanding, but simply because 81 is the most statistically likely set of tokens corresponding to that prompt per its training set.
And yet when it produces these correct results, people ooh and aah over how "smart" it is, how much it has "understood", how "good it is at counting; better than my son!", and when it produces incorrect results people deride it as dumb and so forth, and and all of this, all of this, is pareidolia; it is neither smart in the one case nor dumb in the other, it does not learn in the sense the word is normally used, it does no counting. We're anthropomorphizing an algorithm that is doing nothing like what we imagine it to do, because we mistake the statistical order in its expressions for the presence of a meaning intended by those expressions. It's all projection on our end.
> What you’re really complaining about is creation of art from apparent randomness. Not the AI model alone but monkeys on a typewriter getting something compelling from the AI.
You accuse others of anthrpormorphisizing the tool but you do the same. Art created with Chat GPT is not created by Chat GPT. It is created by a human using a chat GPT. There is no intrinsic limitation on the profundity of art created using chat GPT or other algorithms.
It’s like complaining that paint is stupid. A comment that is largely irrelevant to the artistic merit of paintings.
Sure, in approximately the same way that the CEO of Sunrise is an animator. Pull the other one, it's got bells on.
Yours is an utterly incoherent interpretation; when ChatGPT outputs that there are 12 characters in the word 13, I have not "created the meaning" 12. You're just fixated on this "actually I am le real artist for typing prompts" axe you want to grind, but it has fuck all to do with anything I'm saying.
My statement is simply that the algo’s are a tool. And tools can be used to make good art.
Some of my favorite creative works came from some awful people and others came from algorithms.
I don't care. It does not effect the works or the way in which the works effect me.
All algorithms are made by humans and/or process human input.
And besides, I never said creativity is a requirement for appreciation. I appreciate things in nature regardless of the fact they weren’t the result of creativity.
If asked in good faith (not assuming the answer), I can agree that it's an important question.
For the same reason it ground Kurt Cobain's gears.
"He knows all our pretty songs, and he likes to sing along. But he don't know what it means"
I always thought that was a bit condescending but it applies perfectly to chatgpt.
Why do you think that?
If we had an widely accepted answer, how would the world be different?
(i don't believe intellectual property is a morally legitimate concept, since it comes from a exploration of a pre existing space of ideas (also I am a georgist so I don't believe physical space can be morally owned either))
naturally this strongly held belief can result in sharp words against perceived enemies.
So I currently find myself saying "we must get a handle on this copyright infringement" because equal protection is important, while also believing that copyright protections ought to be much weaker than they are.