I was recently talking with a Daniel Kokotajlo about AI art. It turned out that he and I initially disagreed about ethical questions, but by the end of the conversation, I had somewhat won him over to my position.
I have the vague impression that a lot of people (on the technology side) haven't thought through these issues so much, or (like me) have only recently thought these issues through (as a result of artists making a lot of noise about it!).
So I thought I would write a post. Maybe it will be persuasive to some readers.
Is this the most important conversation to be having about AI?
No. Copyright-adjacent issues with AI art are less important than AI-induced unemployment, which is in turn less important than the big questions about the fate of the human race.
However, it's possible that copyright-adjacent issues around intellectual property and AI will be one of the first major issues thrusting AI into the political sphere, in which case this discussion may help to shape public policy around AI for years to come.
The basic issues.
Large language models such as GPT, and AI image generators such as DALL-E, Imagen, Stable Diffusion, etc etc are (very often) trained on copyrighted works without the permission of the copyright holder.[1] This hasn't proven to be a legal problem, yet, but "legal" doesn't mean "ethical".
When models like GPT and DALL-E started coming out, I recall having the thought: oh, it's nice how these models don't really need to worry about copyright, because (I thought) deep learning turns out to generalize quite well, which means deep-learning-based systems aren't liable to regurgitate copyrighted material.
This turns out to be simply false; these systems are in fact quite liable to reproduce, or very nearly reproduce, copyrighted material when prompted in the right way.[2]
Whether or not copyrighted material is precisely reproduced, or nearly reproduced, or not reproduced at all, there is, in any case, an argument to be made that these AI systems (if/when they charge for use) are turning a profit based on copyrighted material in an illegitimate way.
After all: the purpose of copyright law is, to a very large extent, to preserve the livelihood of intellectual property creators, who would otherwise have limited ability to profit from their own works due to the ease of reproducing it once made. Modern AI systems are threatening this, whether or not they technically violate copyright.
But I want to firmly distinguish between a few different issues:
AI systems training on copyrighted data without the consent of the copyright holder. This is the main issue I will discuss.
AI systems being capable of reproducing copyrighted works exactly or almost exactly. This is a consequence of the first bullet point, plus properties of modern ML systems, plus the absence of safeguards specifically preventing this from happening.[3]
AI systems imitating work in a more general sort of way, such as copying the style of specific artists who never consented to their work being used as training data. This is one of the main reasons to think that training on copyrighted work (without permission) has occurred, in cases where there isn't much public information about what data was used to train an AI. It is also one of the main reasons (I have seen) that artists want these systems to stop training on copyrighted works.[4]
AI putting artists and writers out of work. This is not the main topic of the post, but is an obvious underlying reason why people might be upset.
Artists who take a position against AI art will sometimes describe the situation as follows: AI programmers steal our art, and use it to train AIs, which can then steal our artistic style, and thereby deprive us of business and livelihood (because AI can do it cheaper).
I'm having real trouble finding out about Dall E and copyright infringement. There are several comments about how Dall E can "copy a style" without it being a violation to the artist, but seriously, I'm appalled.
Some defenders of AI art then object, saying the law does not consider it theft, therefore no theft has taken place.
I am not a lawyer, and confess to ignorance about how the law currently treats AI or the likely outcome of court cases about it.
However, it seems clear to me that the current legal system is an attempt to codify reasonable rules in the absence of significant AI technology. The fact(?) that it's not legally considered theft doesn't mean it's not morally theft in a significant sense.
It seems to me like we're at a point where it would be very reasonable to have a society-wide conversation about what should and shouldn't be allowed.
It's what humans do.
Human artists "train on copyrighted works" (ie, look at what other artists do and take inspiration from it). Furthermore, "fair use" allows humans to make significant use of copyrighted works, so long as the new work is "transformative" of the copyrighted material (amongst a short list of other fair-use conditions, including educational use).
Shouldn't we just treat AI the same way? So isn't "training on copyrighted material" fine?
It seems to me that the only thing that seems possible is to treat it like a human that took inspiration from many sources. In the vast majority of cases, the sources of the artwork are not obvious to any viewer (and the algorithm cannot tell you one). Moreover, any given created piece is really the combination of the millions of pieces of the art that the AI has seen, just like how a human takes inspiration from all of the pieces that it has seen. So it seems most similar to the human category, not the simple manipulations (because it isn’t a simple manipulation of any given image or set of images).
Again, I would argue that this is a new situation which very well may call for different norms from the human case. Here are a few differences which we might consider relevant:
Human artist learning from (copyrighted) works:
AI learning from (copyrighted) works:
Not very output-scalable. One human can only do so much work.
Very very output-scalable. Once you've trained a network, producing work is relatively inexpensive. One AI can disrupt the whole market. This is much less of a "level playing field".
Not very input-scalable. One human can only see so much media.
Much more input-scalable. Modern systems are trained on a significant fraction of human-produced media. Again, less of a "level playing field".
Humans form rich generalizations from a small number of examples.
Deep learning systems require huge amounts of data to approach human-level generalizations. This indicates, to an extent, that what's learned from a single example is "shallow". Perhaps this could be seen as closer to plagiarism.
Humans can understand and avoid the idea of copyright violation, and are often cautious to "not steal ideas" even beyond the legal requirements. With some notable exceptions, humans are really trying to create unique works.
Most current AI systems have no safeguards with respect to copyright violations, and certainly don't have the human idea of "not stealing ideas". Indeed, to a large extent, these systems are being trained to mimic their input data as closely as possible.
It's a human, gosh darn it!
It's not a human, gosh darn it! As anthropocentric as the idea may be, it's pretty standard for the law to treat humans differently.
My opinion would be that this calls for a civilization-wide discussion of what the new norms should be.
There's no precedent for calling this immoral.
"Sure", you say,[6] "There's no precedent for AI creativity at the level we're now seeing. But I'm afraid that argument cuts both ways. You can't call modern training methods 'unethical' out of the blue. If there had been previous illustrations of this kind of dilemma in science fiction, for example, with a clear consensus amongst sci-fi authors that training AIs on copyrighted works would be considered unethical, fine. But prior to current complaints, there was no such consensus against these techniques! Artists are clearly making up new ethical rules because they are upset about losing jobs."
Counter #1: But I vaguely felt like there was a consensus on this?!
You could easily accuse me of hindsight bias and/or constructed memory, but as I've already mentioned, I recall assuming that OpenAI and other companies had done their due diligence to make sure that they weren't stepping over the line.
I imagine a lot of other AI-oriented grad students have thought about trying to train image-generation stuff at one point or another in their career. I certainly did. I have the impression that, say, 2014-me included in such plans steps such as "obtain permission from the artists, or otherwise, seek out training material that has fallen out of copyright."
This is definitely more like "academic caution" than "legal caution"; but it's standard practice in academia to make attribution clear, just as it is in art. It seems like just a mistake to think that caution about proper attribution should go away when those two worlds cross over.
For example, I think there's a clear academic consensus that you should obtain permission (and properly attribute) if you reproduce someone else's figure in your paper. It doesn't make a difference whether it's publicly available on the web.
It's not a logical deduction or anything, but it seems to me like natural academic caution about attribution extends to the point where you ask copyright holders before using copyrighted data to train an AI.
I also seem to recall a very early writing-assistance tool based on Markov chains (I'm not claiming it was commercially successful or anything), which advertised, as an explicit feature that it had a filter to make absolutely sure that it would not auto-suggest sections from copyrighted works. This isn't a precedent for "don't train on copyrighted works without permission", but it is a precedent for "be cautious around copyright", and in particular "put precautions in place to make sure your AI doesn't reproduce copyrighted work".
Counter #2: There's a clear moral consensus about user data.
Another argument which Daniel Kokatajlo pointed out to me is that in recent years, there has been a growing consensus that there's something skeezy about harvesting user data and using it for things in general, especially without transparency about what's happening.
Harvesting data to train AI, without consent from the original creators, seems like it falls under this.
The Case For Dialogue?
In discussions like this, it's easy for one side to demonize or dismiss the other side. I think a lot of the problems here are arising because programmers weren't really thinking of artists at all when they made certain decisions. (Of course, this is only a guess.)
I could write a long rant about my exact critique of that discussion, but I guess it would not be very interesting to read.
Basically, I think it could be done better. However, I worry that if I had a super-public conversation with an artist like this, my personal views would inevitably get attributed to MIRI, and this doesn't seem so good. I think other people who work for prominent organizations are in a similar position.
So I guess I'm saying: consider whether it might be worth a little of your time to reach out to artists, or (if you're an artist) reach out to AI programmers, or otherwise facilitate such conversations?
I think it's moderately plausible that this becomes an important issue in another election cycle or two, and a little plausible that conversations which take place now could help.
I'm not sure exactly which systems were and were not trained on copyrighted material; and in some cases, I think the information is not publicly available. The fact that most/all modern deep-learning image-generation tools I am aware of can copy the styles of a broad variety of specific artists when asked seems like significant evidence that most/all of these systems have been trained on copyrighted material.
But at least we know that Stable Diffusion has been, since its data-set is public.
I initially thought that modern ML (meaning, very very large transformer networks) was safe from this kind of risk because it showed an ability to generalize very well, and be very creative when output was generated by random sampling.
However, it turns out that modern ML memorizes its data quite well, meaning that it achieves extremely low loss when the same work is shown to it again during training. This means it's possible for it to generate stuff directly from its training data, just by sampling.
On the pro-AI-art side, I've seen the argument made that modern ML can't be memorizing its training data, since the size of the neural network (in bytes) is far far smaller than the size of the data-set. But this seems to be wrong.
Obviously, it's possible to compress the training data a lot. Obviously, it's possible for the network to memorize some things but not all.
Being able to reproduce the style of a specific artist hurts the livelihood of that artist in ways that AI art in general does not. It allows scammers to pretend to be that artist, for example on social media websites. It also allows companies to produce products which use the style, where previously they would be forced to pay the original artist (or a good human imitator, which can be harder to find and might not save any money).
Deep learning systems require huge amounts of data to approach human-level generalizations. This indicates, to an extent, that what's learned from a single example is "shallow". Perhaps this could be seen as closer to plagiarism.
The lawsuit against Stable Diffusion argues that SD works by amassing a huge library of images that the system then interpolates between in order to generate the desired kinds of images, but struggles to create the kinds of image combinations that don't appear in the training data and thus can't be interpolated between. Some of my friends have also remarked on this, e.g. that there are many contexts where it's a struggle to get the system to draw women in a non-sexualized way. (See also Scott Alexander on the way that DALL-E conflates style and content.)
This is then different from the kind of learning that a human artist does - humans don't only store a huge library of reference photos in their mind and interpolate between them, but they actually get a conceptual understanding of the world as well. Because of that, they could easily draw pictures even of things they've never seen before ("a dog wearing a baseball cap while eating ice cream" is the example used in the complaint). In contrast, systems like Stable Diffusion are limited to only being able to draw things that are a sufficiently close match to images they've already seen. In that sense, a human artist who draws the kind of a picture that would otherwise not have existed in SD's training set is much more directly enabling the system to draw those kinds of pictures, than they would be enabling another human artist to do the same. (Or so the argument goes.)
Ho showed how a latent image could be interpolated—meaning, blended mathematically—to produce new derivative images. Rather than combine two images pixel by pixel—which gives unappealing results—Ho showed how Training Images can be stored in the diffusion model as latent images and then interpolated as a new latent image. This interpolated latent image can then be converted back into a standard pixel-based image.
The diagram below, taken from Ho’s paper, shows how this process works, and demonstrates the difference in results between interpolating pixels and interpolating latent images.
In the diagram, two photos are being blended: the photo on the left labeled “Source x0,” and the photo on the right labeled “Source x'0.”
The image in the red frame has been interpolated pixel by pixel, and is thus labeled “pixel-space interpolation.” This pixel-space interpolation simply looks like two translucent face images stacked on top of each other, not a single convincing face.
The image in the green frame, labeled “denoised interpolation”, has been generated differently. In that case, the two source images have been converted into latent images (illustrated by the crooked black arrows pointing upward toward the label “Diffused source”). Once these latent images have been interpolated (represented by the green dotted line), the newly interpolated latent image (represented by the smaller green dot) has been reconstructed into pixels (a process represented by the crooked green arrow pointing downward to a larger green dot). This process yields the image in the green frame. Compared to the pixel-space interpolation, the difference is apparent: the denoised blended interpolation looks like a single convincing human face, not an overlay or combination of images of two faces. [...]
Despite the difference in results, these two modes of interpolation are equivalent: they both generate derivative works from the source images. In the pixel-space interpolation (the red-framed image), the source images themselves are being directly interpolated to make a derivative image. In the denoised interpolation (the green-framed image), (1) the source images are being converted to latent images, which are lossy-compressed copies; (2) those latent images are being interpolated to make a derivative latent image; and then (3) this derivative latent image is decompressed back into a pixel-based image.
In April 2022, the diffusion technique was further improved by a team of researchers led by Robin Rombach at Ludwig Maximilian University of Munich. These ideas were introduced in his paper “High-Resolution Image Synthesis with Latent Diffusion Models.”
Rombach is also employed by Stability as one of the primary developers of Stable Diffusion, which is a software implementation of the ideas in his paper.
Rombach’s diffusion technique offered one key improvement over previous efforts. Rombach devised a way to supplement the denoising process by using extra information, so that latent images could be interpolated in more complex ways. This process is called conditioning. The most common tool for conditioning is short text descriptions, previously introduced as Text Prompts, that might describe elements of the image, e.g.—“a dog wearing a baseball cap while eating ice cream”. This metric uses Text Prompts as conditioning data to select latent images that are already associated with text captions indicating they contain “dog,” “baseball cap,” and “ice cream.” The text captions are part of the Training Images, and were scraped from the websites where the images themselves were found.
The resulting image is necessarily a derivative work, because it is generated exclusively from a combination of the conditioning data and the latent images, all of which are copies of copyrighted images. It is, in short, a 21st-century collage tool.
The result of this conditioning process may or may not be a satisfying or accurate depiction of the Text Prompt. Below is an example of output images from Stable Diffusion (via the DreamStudio app) using this Text Prompt—“a dog wearing a baseball cap while eating ice cream”. All these dogs in the resulting images seem to be wearing baseball hats. Only the one in the lower left seems to be eating ice cream. The two on the right seem to be eating meat, not ice cream.
In general, none of the Stable Diffusion output images provided in response to a particular Text Prompt is likely to be a close match for any specific image in the training data. This stands to reason: the use of conditioning data to interpolate multiple latent images means that the resulting hybrid image will not look exactly like any of the Training Images that have been copied into those latent images.
But it is also true that the only thing a latent-diffusion system can do is interpolate latent images into hybrid images. There is no other source of visual information entering the system.
Every output image from the system is derived exclusively from the latent images, which are copies of copyrighted images. For these reasons, every hybrid image is necessarily a derivative work.
A latent-diffusion system can never achieve a broader human-like understanding of terms like “dog,” “baseball hat,” or “ice cream.” Hence, the use of the term “artificial intelligence” in this context is inaccurate.
A latent-diffusion system can only copy from latent images that are tagged with those terms. The system struggles with a Text Prompt like “a dog wearing a baseball cap while eating ice cream” because, though there are many photos of dogs, baseball caps, and ice cream among the Training Images (and the latent images derived from them) there are unlikely to be any Training Images that combine all three.
A human artist could illustrate this combination of items with ease. But a latentdiffusion system cannot because it can never exceed the limitations of its Training Images.
In practice, the quality of the latent-diffusion images depends entirely on the breadth and quality of the Training Images used to generate the latent images. If that weren’t true, then it wouldn’t matter where Stable Diffusion (or any other AI-Image Product) got its Training Images.
In actuality, the provenance of an AI-Image-Product’s Training Images matters very much. According to Emad Mostaque, CEO of Stability, Stable Diffusion has “compress[ed] the knowledge of over 100 terabytes of images.” Though the rapid success of Stable Diffusion has been partly reliant on a great leap forward in computer science, it has been even more reliant on a great leap forward in appropriating copyrighted images.
What's amusing is before this case ever even sees a trial, the above limitations may be overcome. Feedback from a system that checks the output image actually satisfies the prompt and that humans have the correct number of fingers for instance.
Interestingly i believe this is a limitation that one of the newest (as yet unreleased) diffusion models has overcome, called DeepFloyd; a number of examples have been teased already, such as the following Corgi sitting in a sushi doghouse:
"After all: the purpose of copyright law is, to a very large extent, to preserve the livelihood of intellectual property creators, who would otherwise have limited ability to profit from their own works due to the ease of reproducing it once made. Modern AI systems are threatening this, whether or not they technically violate copyright."
While it's probably true that copyright/patent/IP law generally in effect helps "preserve the livelihood of intellectual property creators," it's a mistake IMO to see this as more than merely instrumental in preserving incentives for more art/inventions/technology which, but for a temporary monopoly (IP protections), would be financially unprofitable to create. Additionally, this view ignores art consumers, who out-number artists by several orders of magnitude. It seems unfair to orient so much of the discussion of AI art's effects on the smaller group of people who currently create art.
IMO the key questions (both morally & legally) should fall into two camps:
Value Creation
I.e, whether, in a regime where to training algos on copyrighted works is permissible, there are
higher volumes of art to consume/appreciate
"better"/more aesthetically pleasing art
than in a regime where people can only train AI art/inventions on public domain & proprietary art/inventions.
No. 2 seems pretty clearly true, but I'm struggling to articulate why. No 1. Seems somewhat conditional on No 2, since I suspect there would be less art created if the AI art tools create "worse" art.
Enforcement Costs
I.e whether - conditional on copyright "infringing algos yielding net societal equal or lower terminal art//innovation volume and/or equal/diminished quality - the detection and enforcement costs of techniques to stop the creation of art from algorithms trained on copyrighted works are sufficiently low.
I doubt there's a lot of societal value in creating an expensive cottage industry of copyright inspectors whose end output degrades the aggregate quality of humanity's art-stock. I don't have priors for the costs of such an enforcement mechanism, but IP lawyers seem expensive & regulatory orgs can get bloated pretty easily.
While it's probably true that copyright/patent/IP law generally in effect helps "preserve the livelihood of intellectual property creators," it's a mistake IMO to see this as more than merely instrumental in preserving incentives for more art/inventions/technology which, but for a temporary monopoly (IP protections), would be financially unprofitable to create. Additionally, this view ignores art consumers, who out-number artists by several orders of magnitude. It seems unfair to orient so much of the discussion of AI art's effects on the smaller group of people who currently create art.
I think you've got this precisely backwards. The concept of laws as such only makes sense in a deontological framework where the fruits of intellectual labor belong to the individual who produced them. Otherwise instead of complicated rules about temporary monopolies and intellectual property, the government would just allow any use which could be proven in court to be net positive in utility, regardless of the wishes of the original creator.
Whether or not you think this is a bad idea, I think it clear that society at large doesn't agree with the framework you've proposed for evaluating IP and copyright.
Actually you got it backwards. The so called intellectual property doesn’t have typical attributes of property:
– exclusivity: if I take it from you, you don’t have it anymore
– enforceability: it’s not trivial to even find out my “art was stolen”
– independence: I can violate your IP by accident even if I never seen any of your works (typical for patents), this can’t happen with proper property
– clear definition: you usually don’t need courts to decide whether I actually took your car or not.
Besides that, IP is in direct conflict with proper property rights (right to use your own property freely).
However, having IP is a practical way of overcoming the black passenger problem. But that’s the reason it was created in the first place. That’s the reason it actually expires after some time and works become a part of “public domain”. (Can you imagine a car becoming a part of public domain? See the difference?)
Now, even the US constitution is aware of this and explicitly states “progress of science and arts” as the only lawful reason to enact copyright.
[The Congress shall have power] “To promote the progress of science and useful arts, by securing for limited times to authors and inventors the exclusive right to their respective writings and discoveries.
"After all: the purpose of copyright law is, to a very large extent, to preserve the livelihood of intellectual property creators, who would otherwise have limited ability to profit from their own works due to the ease of reproducing it once made. Modern AI systems are threatening this, whether or not they technically violate copyright."
Yes, this is 100% backwards. The purpose of copyright law is to incetivise the production of art so that consumers of art can benefit from it. It incidentally protects artists livelihoods, but that is absolutely not it's main purpose.
We only want to protect the livelihood of artists because humans enjoy consuming art- the consumption is the ultimate point. We don't have laws protecting the livelihood of people who throw porridge at brick walls because we don't value that activity. We also don't have laws protecting the livelihood of people who read novels, because while lots of people enjoy doing that, other people don't value the activity.
If we can get art produced without humans invovled, that is 100% a win for society. In the short term it puts a few people out of work, which is unfortunate, but short-lived. The fact that AI art is vastly more efficiently-produced than human art is a good thing, that we should be embracing.
I think you're mainly correct, but it's a bit of both. We have laws and subsidies protecting the livelihoods of farmers. The way democracy works, the winning coalition will cater to its constituents by passing laws which benefit those folks. So plausibly, the winning coalitions which passed/protected intellectual property laws included some support by artists, too. (This is especially plausible if you think about how Disney influences copyright law.)
Given the current reaction to AI art, I think it's plausible (but very uncertain) that enough people would side with artists here to democratically protect artists now/soon. People enjoy consuming art, but doing so also creates some degree of emotional connection to the artists themselves (a parasocial relationship).
While it's probably true that copyright/patent/IP law generally in effect helps "preserve the livelihood of intellectual property creators," it's a mistake IMO to see this as more than merely instrumental in preserving incentives for more art/inventions/technology which, but for a temporary monopoly (IP protections), would be financially unprofitable to create.
I'm not sure what you're saying here! My implication was that we should view the law as instrumental rather than terminally valuing the law as it currently stands. I don't know much about the law, but I also have the impression that judges will think about it this way when considering how to respond to this new situation.
Additionally, this view ignores art consumers, who out-number artists by several orders of magnitude. It seems unfair to orient so much of the discussion of AI art's effects on the smaller group of people who currently create art.
True! In the past, protecting the profitability of artists this way was also (for the most part) to the benefit of consumers, since profitability of art determined how much was created and mass-distributed. Especially before the internet.
No. 2 seems pretty clearly true, but I'm struggling to articulate why. No 1. Seems somewhat conditional on No 2, since I suspect there would be less art created if the AI art tools create "worse" art.
AI art generally seems like a lot of #1 and only a little #2, right now. Obviously the quality will keep getting better.
If training on copyrighted work was outlawed tomorrow, then I think we would see less AI art in the very short term (so negative impact to #1 temporarily), and in the medium term, less human artists out of a job (so, somewhat temporary positive impact to #2).
In the longer term, I think it's not going to matter very much, since the technology will find ways to improve one way or another.
Enforcement Costs
I personally imagine enforcement costs will be low, because training these systems requires large amounts of money and is accomplished by a relatively small number of orgs which will mostly be self-policing once the legal situation is clear (because the risk of investing that much money, and then having a court tell you to throw the result away, is going to be mostly unacceptable).
Additionally, this view ignores art consumers, who out-number artists by several orders of magnitude. It seems unfair to orient so much of the discussion of AI art's effects on the smaller group of people who currently create art.
What is the greater framework behind this argument? "Creating art" is one of the most general potentials a human being can realize. With your argument we could justify chopping off every human potential because "there's a greater amount of people who don't care about realizing it".
I think deleting a key human potential (and a shared cultural context) affects the entire society.
It's funny that short-timeline-believers tend not to care much about the topic, as it'll be very minor very soon. And long-timeline-believers think we've got at least a little breathing room to sort it out using slow human processes for social and legal norm adjustment.
I put myself somewhere in between. We probably don't have the 2+ (human) generations it takes to societally absorb a giant change, but it's not really a crisis yet. We haven't seen any significant court case outcomes NOR legislation that needs court testing (I really am looking forward to the Getty case, though).
Artists (and other "creatives") are worried, far more concerned that their future artistic positioning and revenue will be reduced by "unfair" competition, than that their copyright exclusivity for past work will be violated. This seems to me to be the most important aspect: the future of human work-value (especially non-elite work). I think it's surprising a lot of us that the "creative" work seems to be under more attack than the "rote" work (driving, warehouse, etc.). I don't know what the new equilibrium will be, and I can't see any simple solutions.
It's deeply unfortunate that the US no longer has any ability to actually discuss, compromise, and experiment on policy. Culture wars take over too soon, and this prevents any sensible small-steps or even measurement of such changes.
I'm not sure what short timeline bettors you're thinking of here, but I personally think that ai art is pretty much the only form the ai safety problem will ever take. Art is a generative model's paperclip.
Reasonable points, all! I agree that the conflation of legality and morality has warped the discourse around this; in particular the idea of Stable Diffusion and such regurgitating copyrighted imagery strikes me as a red herring, since the ability to do this is as old as the photocopier and legally quite well-understood.
It actually does seem to me, then, that style copying is a bigger problem than straightforward regurgitation, since new images in a style are the thing that you would ordinarily need to go to an artist for; but the biggest problem of all is that fundamentally all art styles are imperfect but pretty good substitutes in the market for all other art styles.
(Most popular of all the art styles-- to judge by a sampling of images online-- is hyperrealism, which is obviously a style that nobody can lay either legal OR moral claim to.)
So i think that if Stability tomorrow came out with a totally unimpeachable version of SD with no copyrighted data of any kind (but with a similarly high quality of output) we would have, essentially, the same set of problems for artists.
I'm confused about how style copying is a new problem. You can trivially find people willing a capable of drawing convincing Disney or specific-anime-studio art, and there's an entire town in China dedicated to making paintings in famous styles. This has existed for a long time and the moral panic is just because now scary computers are doing it.
So i think that if Stability tomorrow came out with a totally unimpeachable version of SD with no copyrighted data of any kind (but with a similarly high quality of output) we would have, essentially, the same set of problems for artists.
I don't think this is true in the short term. Artists are currently dealing with issues like scam social media accounts which copy their style and claim to be the artist. (Not sure how big this is, I only heard about this as a rumor -- but it's something that is now possible, where before you'd only be able to do something like this by re-posting existing works.)
Very well written, thank you! All of the writing about AI-generated art that I've stumbled across has been either one-sentence talking points (e.g. "it's stealing art without artists' permission" or "training an AI model is just like a human looking at past art") or hedgy arguments from news articles ("some artists are concerned that...").
It's refreshing to see a serious, grounded look at the ethics of AI art. I was thinking about writing my own post along the same vein, but this covers most of what I would have touched on (and more).
Video embeds for relevant videos - first the stilted conversation between random frantic ai nerd who is trying to clarify that there's nothing that can be done to stop ai and we better hurry (I agree with him, he didn't make it clear enough how hopeless it is to stop it, though, too many people are like "why not just not?" and don't understand why that's ... nigh on not permitted by physics)
And a couple of related videos I'd recommend, both from the past couple of days. Both are best watched 2x speed with captions imo. Or toss them in whisper and just read the video. It's good research regardless, these are just blog posts, it's just that most blog posts are videos because videos get more normie engagement. Sorry.
(And as usual I try to be a hub of "stuff people should have been aware of already", spider links manually from my userpage or dm me for more links. Basically, holy shit check out IPAM.)
I was recently talking with a Daniel Kokotajlo about AI art. It turned out that he and I initially disagreed about ethical questions, but by the end of the conversation, I had somewhat won him over to my position.
I have the vague impression that a lot of people (on the technology side) haven't thought through these issues so much, or (like me) have only recently thought these issues through (as a result of artists making a lot of noise about it!).
So I thought I would write a post. Maybe it will be persuasive to some readers.
Is this the most important conversation to be having about AI?
No. Copyright-adjacent issues with AI art are less important than AI-induced unemployment, which is in turn less important than the big questions about the fate of the human race.
However, it's possible that copyright-adjacent issues around intellectual property and AI will be one of the first major issues thrusting AI into the political sphere, in which case this discussion may help to shape public policy around AI for years to come.
The basic issues.
Large language models such as GPT, and AI image generators such as DALL-E, Imagen, Stable Diffusion, etc etc are (very often) trained on copyrighted works without the permission of the copyright holder.[1] This hasn't proven to be a legal problem, yet, but "legal" doesn't mean "ethical".
When models like GPT and DALL-E started coming out, I recall having the thought: oh, it's nice how these models don't really need to worry about copyright, because (I thought) deep learning turns out to generalize quite well, which means deep-learning-based systems aren't liable to regurgitate copyrighted material.
This turns out to be simply false; these systems are in fact quite liable to reproduce, or very nearly reproduce, copyrighted material when prompted in the right way.[2]
Whether or not copyrighted material is precisely reproduced, or nearly reproduced, or not reproduced at all, there is, in any case, an argument to be made that these AI systems (if/when they charge for use) are turning a profit based on copyrighted material in an illegitimate way.
After all: the purpose of copyright law is, to a very large extent, to preserve the livelihood of intellectual property creators, who would otherwise have limited ability to profit from their own works due to the ease of reproducing it once made. Modern AI systems are threatening this, whether or not they technically violate copyright.
But I want to firmly distinguish between a few different issues:
Some initial arguments.
It's not illegal.[5]
Artists who take a position against AI art will sometimes describe the situation as follows: AI programmers steal our art, and use it to train AIs, which can then steal our artistic style, and thereby deprive us of business and livelihood (because AI can do it cheaper).
Several months ago, A.R. Stone made a LessWrong comment somewhat along these lines (quoted in part):
Some defenders of AI art then object, saying the law does not consider it theft, therefore no theft has taken place.
I am not a lawyer, and confess to ignorance about how the law currently treats AI or the likely outcome of court cases about it.
However, it seems clear to me that the current legal system is an attempt to codify reasonable rules in the absence of significant AI technology. The fact(?) that it's not legally considered theft doesn't mean it's not morally theft in a significant sense.
It seems to me like we're at a point where it would be very reasonable to have a society-wide conversation about what should and shouldn't be allowed.
It's what humans do.
Human artists "train on copyrighted works" (ie, look at what other artists do and take inspiration from it). Furthermore, "fair use" allows humans to make significant use of copyrighted works, so long as the new work is "transformative" of the copyrighted material (amongst a short list of other fair-use conditions, including educational use).
Shouldn't we just treat AI the same way? So isn't "training on copyrighted material" fine?
In the same thread as the A.R. Stone comment I mentioned earlier, gbear605 makes an argument along these lines (quoted in part):
Again, I would argue that this is a new situation which very well may call for different norms from the human case. Here are a few differences which we might consider relevant:
My opinion would be that this calls for a civilization-wide discussion of what the new norms should be.
There's no precedent for calling this immoral.
"Sure", you say,[6] "There's no precedent for AI creativity at the level we're now seeing. But I'm afraid that argument cuts both ways. You can't call modern training methods 'unethical' out of the blue. If there had been previous illustrations of this kind of dilemma in science fiction, for example, with a clear consensus amongst sci-fi authors that training AIs on copyrighted works would be considered unethical, fine. But prior to current complaints, there was no such consensus against these techniques! Artists are clearly making up new ethical rules because they are upset about losing jobs."
Counter #1: But I vaguely felt like there was a consensus on this?!
You could easily accuse me of hindsight bias and/or constructed memory, but as I've already mentioned, I recall assuming that OpenAI and other companies had done their due diligence to make sure that they weren't stepping over the line.
I imagine a lot of other AI-oriented grad students have thought about trying to train image-generation stuff at one point or another in their career. I certainly did. I have the impression that, say, 2014-me included in such plans steps such as "obtain permission from the artists, or otherwise, seek out training material that has fallen out of copyright."
This is definitely more like "academic caution" than "legal caution"; but it's standard practice in academia to make attribution clear, just as it is in art. It seems like just a mistake to think that caution about proper attribution should go away when those two worlds cross over.
For example, I think there's a clear academic consensus that you should obtain permission (and properly attribute) if you reproduce someone else's figure in your paper. It doesn't make a difference whether it's publicly available on the web.
It's not a logical deduction or anything, but it seems to me like natural academic caution about attribution extends to the point where you ask copyright holders before using copyrighted data to train an AI.
I also seem to recall a very early writing-assistance tool based on Markov chains (I'm not claiming it was commercially successful or anything), which advertised, as an explicit feature that it had a filter to make absolutely sure that it would not auto-suggest sections from copyrighted works. This isn't a precedent for "don't train on copyrighted works without permission", but it is a precedent for "be cautious around copyright", and in particular "put precautions in place to make sure your AI doesn't reproduce copyrighted work".
Counter #2: There's a clear moral consensus about user data.
Another argument which Daniel Kokatajlo pointed out to me is that in recent years, there has been a growing consensus that there's something skeezy about harvesting user data and using it for things in general, especially without transparency about what's happening.
Harvesting data to train AI, without consent from the original creators, seems like it falls under this.
The Case For Dialogue?
In discussions like this, it's easy for one side to demonize or dismiss the other side. I think a lot of the problems here are arising because programmers weren't really thinking of artists at all when they made certain decisions. (Of course, this is only a guess.)
I was really glad to see a dialogue between a San Francisco techie and a prominent YouTube art channel. However, I was also disappointed by some aspects of the conversation.
I could write a long rant about my exact critique of that discussion, but I guess it would not be very interesting to read.
Basically, I think it could be done better. However, I worry that if I had a super-public conversation with an artist like this, my personal views would inevitably get attributed to MIRI, and this doesn't seem so good. I think other people who work for prominent organizations are in a similar position.
So I guess I'm saying: consider whether it might be worth a little of your time to reach out to artists, or (if you're an artist) reach out to AI programmers, or otherwise facilitate such conversations?
I think it's moderately plausible that this becomes an important issue in another election cycle or two, and a little plausible that conversations which take place now could help.
https://en.m.wikipedia.org/wiki/Stable_Diffusion#Usage
I'm not sure exactly which systems were and were not trained on copyrighted material; and in some cases, I think the information is not publicly available. The fact that most/all modern deep-learning image-generation tools I am aware of can copy the styles of a broad variety of specific artists when asked seems like significant evidence that most/all of these systems have been trained on copyrighted material.
But at least we know that Stable Diffusion has been, since its data-set is public.
https://techcrunch.com/2022/12/13/image-generating-ai-can-copy-and-paste-from-training-data-raising-ip-concerns/
I initially thought that modern ML (meaning, very very large transformer networks) was safe from this kind of risk because it showed an ability to generalize very well, and be very creative when output was generated by random sampling.
However, it turns out that modern ML memorizes its data quite well, meaning that it achieves extremely low loss when the same work is shown to it again during training. This means it's possible for it to generate stuff directly from its training data, just by sampling.
On the pro-AI-art side, I've seen the argument made that modern ML can't be memorizing its training data, since the size of the neural network (in bytes) is far far smaller than the size of the data-set. But this seems to be wrong.
Obviously, it's possible to compress the training data a lot. Obviously, it's possible for the network to memorize some things but not all.
But the most persuasive argument is when we re-generate images almost precisely, with only a text prompt.
I'm not sure exactly which systems have safeguards, or lack them. There was discussion of DALL-E
Being able to reproduce the style of a specific artist hurts the livelihood of that artist in ways that AI art in general does not. It allows scammers to pretend to be that artist, for example on social media websites. It also allows companies to produce products which use the style, where previously they would be forced to pay the original artist (or a good human imitator, which can be harder to find and might not save any money).
Of course, the reality is we don't yet know what's legal or illegal, because this hasn't yet been tested in court.
Daniel Kokotajlo made an argument similar to this.