[Riverheart]

“A legal reality where you can only train AI on content you've licensed would be the worst for everybody bar massive companies, legacy artists included.”

Care to elaborate?

Also, saying artists only concern themselves with the legality of art used in AI because of distaste when there are legal cases where their art has been appropriated seems like a bold position to take.

It’s a practice founded on scooping everything up without care for origin or attribution and it’s not like it’s a transparent process. There are people that literally go out of their way to let artists know they’re training on their art and taunt them about it online. Is it unusual they would assume bad faith from those purporting to train their AI legally when participation up till now has either been involuntary or opt out? Rolling out AI features when your customers are artists is tone deaf at best and trolling at worst.

[Workaccount2]

There is no "scooping up", the models aren't massive archives of copied art. People either don't understand how these models work or they purposely misrepresent it (or purposely refuse to understand it).

Showing the model an picture doesn't create a copy of that picture in it's "brain". It moves a bunch of vectors around that captures an "essence" of what the image is. The next image shown from a totally different artist with a totally different style may well move around many of those same vectors again. But suffice to say, there is no copy of the picture anywhere inside of it.

This also why these models hallucinate so much, they are not drawing from a bank of copies, they are working off of a fuzzy memory.

[TeMPOraL]

> People either don't understand how these models work or they purposely misrepresent it (or purposely refuse to understand it).

Not only that, they also assume or pretend that this is obviously violating copyright, when in fact this is a) not clear, and b) pending determination by courts and legislators around the world.

FWIW, I agree with your perspective on training, but I also accept that artists have legitimate moral grounds to complain and try to fight it - so I don't really like to argue about this with them; my pet peeve is on the LLM side of things, where the loudest arguments come from people who are envious and feel entitled, even though they have no personal stake in this.

[Root_Denied]

>Not only that, they also assume or pretend that this is obviously violating copyright, when in fact this is a) not clear, and b) pending determination by courts and legislators around the world.

Legislation always takes time to catch up with tech, that's not new.

The question I'm see being put forth from those with legal and IP backgrounds is about inputs vs. outputs, as in "if you didn't have access to X (which has some form of legal IP protection) as an input, would you be able to get the output of a working model?" The comparison here is with manufacturing where you have assembly of parts made by others into some final product and you would be buying those inputs to create your product output.

The cost of purchasing the required inputs is not being done for AI, which pretty solidly puts AI trained on copyrighted materials in hot water. The fact that it's an imperfect analogy and doesn't really capture the way software development works is irrelevant if the courts end up agreeing with something they can understand as a comparison.

All that being said I don't think the legality is under consideration for any companies building a model - the profit margins are too high to care for now, and catching them at it is potentially difficult.

There's also a tendency for AI advocates to try and say that AI/LLM's are "special" in some way, and to compare their development process to someone "learning" the style of art (or whatever input) that they then internalize and develop into their own style. Personally I think that argument gives a lot of assumed agency to these models that they don't actually have, and weakens the overall legal case.

[Riverheart]

“Not only that, they also assume or pretend that this is obviously violating copyright, when in fact this is a) not clear, and b) pending determination by courts and legislators around the world.”

Uh huh, so much worse than the people that assume or pretend that it’s obviously not infringing and legal. Fortunately I don’t need to wait for a lawyer to form an opinion and neither do those in favor of AI as you might’ve noticed.

You see any of them backing down and waiting for answer from a higher authority?

[TeMPOraL]

> You see any of them backing down and waiting for answer from a higher authority?

Should they? That's generally not how things work in most places. Normally, if something isn't clearly illegal, especially when it's something too new and different for laws to clearly cover, you're free to go ahead and try it; you're not expected to first seek a go-ahead from a court.

[Riverheart]

You just chided people for having strong opinions about AI infringement without a court ruling to back them up but now you’re saying that creating/promoting an entire industry based on a legal grey area is a social norm that you have no strong feelings about. I would have thought the same high bar to speak on copyright for those who believe it infringes would be applied equally to those saying it does not, especially when it financially benefits them. I don’t think we’ll find consensus.

[bawolff]

This is silly. What are you proposing? A coup to ban AI? Because that is the alternative to waiting for legislators and courts.

[Riverheart]

Never proposed a ban, the issue is copyright, use licensed inputs and I could care less.

Pro AI people need to stop behaving like it’s a foregone conclusion that anything they do is right and protected from criticism because, as was pointed out, the legality of what is being done with unlicensed inputs, which is the majority of inputs, is still up for debate.

I’m just calling attention to the double standard being applied in who is allowed to have an opinion on what the legal outcome should be prior to that verdict. Temporal said people shouldn’t “pretend or assume” that lots of AI infringes on other people’s work because the law hasn’t caught up but the same argument applies equally to them (AI proponents) and they have already made up their mind, independent of any legal authority, that using unlicensed inputs is legal.

The difference in our opinions is that if I’m wrong, no harm done, if they’re wrong, lots of harm has already been done.

I’m trying to have a nuanced conversation but this has devolved into some pro/anti AI, all or nothing thing. If you still think I want to ban AI after this wall of text I don’t know what to tell you dude. If I’ve been unclear it’s not for lack of trying.

[bawolff]

But this is hardly limited to AI.

Copyright is full of grey areas and disagreement over its rules happen all the time. AI is not particularly special in that regard, except perhaps in scale.

Generally the way stuff moves forward is somebody tries something, gets sued and either they win or lose and we move forward from that point.

Ultimately "harm" and "legality" are very different things. Something could be legal and harmful - many things are. In this debate i think different groups are harmed depending on which side that "wins".

If you want to have a nuanced debate, the relavent issue is not if the input works are licensed - they obviously are not, but on the following principles:

- de minimis - is the amount of each individual copyrighted work too small to matter.

- is the AI just extracting "factual" information from the works separate from their presentation. After all each individual work only adjusts the model by a couple bytes. Is it less like copying the work or more like writing a book about the artwork that someone could later use to make a similar work (which would not be copyright infringement if a human did it)

- fair use - complicated, but generally the more "transformative" a work is, the more fair use it would be, and AI is extremely transformative. On the other hand it potentially competes commercially with the original work, which usually means less likely to be fair use (and maybe you could have a mixed outcome here, where the AI generators are fine, but using them to sell competing artwork is not, but other uses are ok).

[Ianal]

[jillyboel]

It's unauthorized commercial use. Which part of that is confusing to you?

[rcxdude]

So is google books, and that got ruled as fair use. That it's being used commercially is not a slam dunk case against an argument for fair use.

[ToucanLoucan]

Training data at scale unavoidably taints models with vast amounts of references to the same widespread ideas that appear repeatedly in said data, so because the model has "seen" probably millions of photos of Indiana Jones, if you ask for an image of an archeologist who wears a hat and uses a whip, it's weighted averages are going to lead it to create something extremely similar to Indiana Jones because it has seen Indiana Jones so much. Disintegrating IP into trillions of pieces and then responding to an instruction to create it with something so close to the IP as to barely be distinguishable is still infringement.

The flip-side to that is the truly "original" images where no overt references are present all look kinda similar. If you run vague enough prompts to get something new that won't land you in hot water, you end up with a sort of stock-photo adjacent looking image where the lighting doesn't make sense and is completely unmotivated, the framing is strange, and everything has this over-smoothed, over-tuned "magazine copy editor doesn't understand the concept of restraint" look.

[tpmoney]

> if you ask for an image of an archeologist who wears a hat and uses a whip, it's weighted averages are going to lead it to create something extremely similar to Indiana Jones because it has seen Indiana Jones so much.

If you ask a human artist for an image of "an archeologist who wears a hat and uses a whip" you're also going to get something extremely similar to Indiana Jones unless you explicitly ask for something else. Let's imagine we go to deviantart and ask some folks to draw us some drawing from these prompts:

A blond haired fighter from a fantasy world that wears a green tunic and green pointy cap and used a sword and shield.

A foreboding space villain with all black armor, a cape and full face breathing apparatus that uses a laser sword.

A pudgy plumber in blue overalls and a red cap of Italian descent

I don't know about you but I would expect with nothing more than that, most of the time you're going to get something very close to Link, Darth Vader and Mario. Link might be the one with the best chance to get something different just because the number of publicly known images of "fantasy world heroes" is much more diverse than the set of "black armored space samurai" and "Italian plumbers"

> Disintegrating IP into trillions of pieces and then responding to an instruction to create it with something so close to the IP as to barely be distinguishable is still infringement.

But it's the person that causes the creation of the infringing material that is responsible for the infringement, not the machine or device itself. A xerox machine is a machine that disintegrates IP into trillions of pieces and then responds to instructions to duplicate that IP almost exactly (or to the best of its abilities). And when that functionality was challenged, the courts rightfully found that a xerox machine in and of itself, regardless of its capability to be used for infringement is not in and of itself infringing.

[ToucanLoucan]

> But it's the person that causes the creation of the infringing material that is responsible for the infringement, not the machine or device itself.

That's simply not good enough. This is not merely a machine that can be misused if desired by a bad actor, this is a machine that specializes in infringement. It's a machine which is internally biased, by the nature of how it works, towards infringement, because it is inherently "copying:" It is copying the weighted averages of millions perhaps billions of training images, many of which depict similar things. No, it doesn't explicitly copy one Indiana Jones image or another: It copies a shit ton of Indiana Jones images, mushed together into a "new" image from a technical perspective, but will inherit all the most prominent features from all of those images, and thus: it remains a copy.

And if you want to disagree with this point, it'd be most persuasive then to explain why, if this is not the case, AI images regularly end up infringing on various aspects of various popular artworks, like characters, styles, intellectual properties, when those things are not being requested by the prompt.

> If you ask a human artist for an image of "an archeologist who wears a hat and uses a whip" you're also going to get something extremely similar to Indiana Jones unless you explicitly ask for something else.

No, you aren't, because an artist is a person that doesn't want to suffer legal consequences for drawing something owned by someone else. Unless you specifically commission "Indiana Jones fanart" I in fact, highly doubt you'll get something like him because an artist will want to use this work to promote their work to others, and unless you are driven to exist in the copyright gray area of fan created works, which is inherently legally dicey, you wouldn't do that.

[tpmoney]

> This is not merely a machine that can be misused if desired by a bad actor, this is a machine that specializes in infringement.

So is a xerox machine. It's whole purpose is to make copies of things whatever you put into it with no regard to whether you have a license to make that copy. Likewise with the record capability on your VCR. Sure you could hook it up to a cam corder and transfer your home movie from a Super-8 to a VHS with your VCR (or like one I used to own, it might even have a camera accessory and port that you could hook a camera up to directly) and yet, I would wager most recordings on most VCRs were to commit copyright infringement. Bit-torrent specializes in facilitating copyright infringement, no matter how many Linux ISOs you download with it. CD ripping software and DeCSS is explicitly about copyright infringement. And let's be real, while MAME is a phenomenal piece of software that has done an amazing job of documenting legacy hardware and its quirks, the entire emulation scene as a whole is built on copyright infringement, and I would wager to a rounding error none of the folks that write MAME emulators have a license to copy the ROMs that they use to do that.

But in all of these cases, the fact that it can (and even usually is) used for copyright infringement is not in and of itself a reason to restrict or ban the technology.

> And if you want to disagree with this point, it'd be most persuasive then to explain why, if this is not the case, AI images regularly end up infringing on various aspects of various popular artworks, like characters, styles, intellectual properties, when those things are not being requested by the prompt.

Well for starters, I'd like to clarify to axioms:

1) "characters" as a subset of "intellectual properties"

2) "style" is not something you can copyright or infringe under US law. It can be part of a trademark or a design patent, and certainly you can commit fraud if you represent something in someone else's style as being a genuine item from that person, but style itself is not protected and I don't think it should be.

So then to answer the question, I would argue that AI images don't "regularly end up infringing on ... intellectual properties, when those things are not being requested by the prompt". I've generated quite a few AI images myself in exploring the various products out there and not a one of them has generated an infringing work, because none of my prompts have asked it to generate an infringing work. It is certainly possible that a given model with a sufficiently limited training set for a given set of words might be likely to generate an infringing image on a prompt, and that's because with a limited set of options to draw from, the prompt is inherently asking for an infringing image no matter how much you try to scrape the serial numbers off. That is, if I ask for an image of "two Italian plumbers who are brothers and battle turtles", everyone knows what that prompt is asking for. There's not a lot of reference options for that particular set of requirements and so it is more likely to generate an infringing image. It's also partly a function of the current goals of the models. As it stands, for the most part we want a model that takes a vague description and gives us something that matches our imagined output. Give that description to most people and they're going to envision the Mario Brothers, so a "good" image generation model is one that will generate a "Mario Brothers" inspired (or infringing) image.

As the technology improves and we get better about producing models that can take new paths without also generating body horror results, and as the users start wanting models that are more creative, we'll begin to see models that can respond to even that limited training set and generate something more unique and less likely to be infringing.

> No, you aren't, because an artist is a person that doesn't want to suffer legal consequences for drawing something owned by someone else.

Sorry, I think you're wrong. If you commission it for money from someone with enough potential visibility, you might encounter people who go out of their way to avoid anything that could be construed as Indiana Jones, but I bet even then you'd get more "Indiana Jones with the serial numbers filed off" images than not.

But if you just asked random artists to draw that prompt, you're going to get an artists rendition of Indiana Jones. It's clear thats what you want from the prompt and that's the single and sole cultural creative reference for that prompt. Though I suppose you and I are going to have to agree to disagree on what people will do unless you're feeling like actually asking a bunch of artist on Fiver to draw the prompt for you.

And realistically what do you expect them to draw when you make that request? When that article showed up with the headline, EVERYONE reading the headline knew the article was talking about an AI generating Indiana Jones. Why did everyone know that? Because of the limited reference for that prompt that exists. "Archeologist that wears a hat and uses a whip" describes very uniquely a single character to almost every single person.

There's a reason no one is writing articles about AIs ripping off Studio Ghibli by showing the output from the prompt "raccoon with giant testicles." No one writes articles talking about how the AI spontaneously generated Garfield knockoffs when prompted to draw an "orange stripped cat". There's no articles about AIs churning out truckloads of Superman images when someone asks for "super hero". And those articles don't exist because there's enough variations on those themes out there, enough different combinations of those words to describe enough different combinations of images and things that those words don't instantly conjure the same image and character for everyone. And so it goes for the AI too. Those prompts don't ask specifically for infringing art so they don't generally generate infringing art.

[Riverheart]

You know why we put up with copyrighted info in the human brain right? Because those are human beings, it’s unavoidable. This? Avoidable.

Also, the model isn’t a human brain. Nobody has invented a human brain.

And the model might not infringe if its inputs are licensed but that doesn’t seem to be the case for most and it’s not clearly transparent they don’t. If the inputs are bad, the intent of the user is meaningless. I can ask for a generic super hero and not mean to get superman but if I do I can’t blame that on myself, I had no role in it, heck even the model doesn’t know what it’s doing, it’s just a function. If I Xerox Superman my intent is clear.

[tpmoney]

> You know why we put up with copyrighted info in the human brain right? Because those are human beings, it’s unavoidable.

I would hope we put up with it because "copyright" is only useful to us insofar as it advances good things that we want in our society. I certainly don't want to live in a world where if we could forcibly remove copyrighted information from human brains as soon as the "license" expired that we would do so. That seems like a dystopian hell worse than even the worst possible predictions of AI's detractors.

> I can ask for a generic super hero and not mean to get superman but if I do I can’t blame that on myself, I had no role in it, heck even the model doesn’t know what it’s doing, it’s just a function.

And if you turn around and discard that output and ask for something else, then no harm has been caused. Just like when artists trace other artists work for practice, no harm is caused and while it might be copyright infringement in a "literal meaning of the words" it's also not something that as a society we consider meaningfully infringing. If on the other hand, said budding artist started selling copies of those traces, or making video games using assets scanned from those traces, then we do consider it infringement worth worrying about.

> If I Xerox Superman my intent is clear.

Is it? If you have a broken xerox machine and you think you have it fixed, grab the nearest papers you can find and as a result of testing the machine xerox Superman, what is your intent? I don't think it was to commit copyright infringement, even if again in the "literal meaning of the words" sense you absolutely did.

[Riverheart]

I’m saying that retaining information is a natural, accepted part of being human and operating in society. Don’t know why it needed to be turned into an Orwell sequel.

[Riverheart]

The collection of the training data is the “scooping up” I mentioned. I assume you acknowledge the training data doesn’t spontaneously burst out of the aether?

As for the model, it’s still creating deterministic, derivative works based off its inputs and the only thing that makes it random is the seed so it being a database of vectors is irrelevant.

[rcxdude]

deterministic is neither here nor there for copyright infringement. a hash of an image is not infringing, and a slightly noisy version of it is.

[Riverheart]

Nobody is trying to copyright an image hash and determinism matters because it’s why the outputs are derivative rather than inspired.

[bawolff]

That is not how copyright works. "Inspired" works can still be derrivative. In the US, entirely deterministic works are not considered derrivative works as they aren't considered new creative works (if anything they are considered the same as the original). See https://en.wikipedia.org/wiki/Bridgeman_Art_Library_v._Corel...

[Riverheart]

“In the US, entirely deterministic works are not considered derrivative works as they aren't considered new creative works (if anything they are considered the same as the original)”

Okay, so if the inputs to the model are my artwork to replicate my style, is the output copyrightable by you? You just said deterministic works aren’t derivative, they’re considered the same as the original. That’s not anything I’ve heard AI proponents claim and the outputs are more original than a 1 to 1 photocopy but I assume like the case you linked to that the answer will be, no, you can’t copyright.