AI images are art, not theft

Far more for my own benefit than anyone else, I want to summarise my perspective from a recent and fruitless discussion. If that thread is anything to go by, then I suspect readers will already agree or disagree with my position, and my arguments here won't persuade anyone either way. Even so, I thought I should set it out, if to no other purpose than to get it out of my system*.

* For the record, in this case I respect my opponent very much, and they absolutely qualify as someone who I still view as credible despite disagreements.

The issue began as whether AI-generated images constitute art. The allegation is that they're not, which spiralled into the accusation that they're actually theft. I here refute both of these claims. There's some overlap between them, but I think I can separate them enough to cover them as different issues.

1) Yes, It's Art

A rigorous, all-encompassing definition of art is well beyond the scope of this post, and arguably impossible given its subjective nature. In fact even fully defining theft might be a bit of a challenge (as we'll see), but I think certain working definitions will do well enough for my purposes here.

As I detailed in the art-meets-science solarigraphy workshop last year :

How you visualise your data really does affect how you interpret it, not just aesthetically but also scientifically : some information is far easier to discern using one technique than another. I think none has a claim to be more valid or true than another, but each can be more appropriate for tackling different questions.

Nobody, including the many professional artists present, raised any issue with this. To me art involves choices, the creation of material for emotional self-expression. If you decide to dance or sing or paint or whatever, you've created something which expresses emotion, and you choose how to communicate your emotions to others. It doesn't really matter what your medium is, and it certainly doesn't matter if it's any good or not : bad art is still art. I would unhesitatingly accept that most AI-images are rubbish (with an annoying, hard-to-define "samey" look), but then so is most conventional art and indeed most content in any field of endeavour.

So what's different about AI-generated images with regards to choice and expression ? Nothing significant. When you write a prompt, you choose what to express. You make the further choice of selecting which of the generated images to share – which one either most closely meets or improves upon (or otherwise modifies in some beneficial way) your expectations. In some systems you even get at least a measure of direct control of the resulting image by using simple drawings to modify the result or generate the initial image. In many others you can refine the resulting image by continuing to respond via the prompt (and always you can simply reject all the output images, write a different prompt and begin from scratch if the output wasn't what you wanted). Just like all art, including writing, there is a measure of both accident and deliberate choice at work.

The counter-argument is that because the user does not (always) have any direct control over the image, then they are not the artist. They are creative only in regards to the prompt, the thinking goes, not the image itself. After all, "causing to be" is not the same as direct, deliberate creation.

But this is to woefully misunderstand the entire creative process. The process of accident and self-reflection is vital, not incidental. Tolkien, for example :

I met a lot of things on the way that astonished me. Tom Bombadil I knew already; but I had never been to Bree. Strider sitting in the corner at the inn was a shock, and I had no more idea who he was than had Frodo. The Mines of Moria had been a mere name; and of Lothlorien no word had reached my mortal ears till I came there. Far away I knew there were the Horselords on the confines of an ancient Kingdom of Men, but Fangorn Forest was an unforeseen adventure. I had never heard of the House of Eorl nor of the Stewards of Gondor. Most disquieting of all, Saruman had never been revealed to me, and I was as mystified as Frodo at Gandalf's failure to appear on September 22nd.

This isn't rhetoric or hyperbole. The vast corpus that is the History of Middle Earth is full on almost every page of accidents in which Tolkien really does "discover", in a sense, what happened, or at least what he wanted to happen : until he set things down on the page, he didn't know for sure (and very often not even then).

Now it's true that you have less direct control in most cases of AI image-generation : the author doesn't move pixels around on the screen. But the analogy of commissioning a piece, claiming that those who pay the piper don't create the tune, doesn't work. I've been on the receiving-payment end of this myself many times and I know from direct first-hand experience that the the client absolutely can have creative control, as stipulated by the contract. "Causing to be" is indeed not the same as "create", this is true, but they're not mutually exclusive either. Without a client paying me money, some works I never would have created at all. Without them further directing me to "move this over there, make that bigger, make that bit more orange", the piece would have ended up quite different from the final result. Claiming that the client had no creative hand in the image, just because they themselves didn't move the pixels on the screen, seems to me to be quite plainly silly.

With AI, the user modifies the image with the only tool available to them : the prompt. They can refine this to the nth degree, just as with commissioning a piece. What's more, the deliberate choice to surrender some level of control is absolutely, unequivocally, unarguably, a valid artistic choice – again see the solarigraphy workshop post. Sometimes this is the very point of art, especially experimental art. Choosing not to choose is still a choice. Writing with a squiggly pen, or using a spirograph (if you arrange its patterns rather than just making one pattern and leaving it at that) : yep, the user is still expressing themselves. AI is just a tool, and the fact it operates differently and can carry out so much more of the process than a paintbrush is the very thing that makes it artistically interesting.

The claim seems to be that in this case, neither the user nor the AI is artistically creating the image, that the whole thing isn't even a creative process at all but an outright sham. I look aghast at this, with the idea that an image can be created without any sort of creativity at all just being incomprehensibly silly. It doesn't have to be the human, emotion-driven creativity, fair enough, but the very fact that something gets created that didn't exist before is self-evidently a creative act. It simply cannot be anything else.

The argument goes that AI doesn't adapt, interpret, or draw on inspiration. It simply regurgitates stolen content, doing nothing more than rearranging stolen data.

I think this is in no sense true. There simply must be adaptation in the sense of transforming text to an image : by its very nature, text is not the same as an image, and the one must be adapted into the other. And there is demonstrably interpretation in that many (sometimes radically) different images can result from the same prompt, just as human artists would have to interpret a request. And there is inspiration in that the models do not simply rearrange pixels, but extract generalised trends to apply them in specific and novel situations. This moves us to the second charge of theft and plagiarism.

2) No, It's Not Stolen

There is at least a legitimate question here as to where creativity ends if it's not strictly limited to the person directly creating the work. Is the guy handing an actor a sandwich responsible for creating Star Wars ? Nah, but what about George Lucas, who wrote and directed, but didn't act in any of the parts or directly create any of the sets or props* ? Clearly there are grey areas and I don't propose a full, generalised answer.

* Did he ? I dunno, but plenty of directors don't, so the specific example doesn't matter.

It's also true that patrons don't always have direct creative control : they can do, but this isn't necessary. Great artists of yesteryear flourished because people paid them money to get on with doing what they were good at without micromanaging them. But when this does happen, when somebody directs the precise details of how an image is constructed – not only what aspects it has but where, their style, arrangement and pose – how can they not be a part of the process ?

Clearly, the guy with the paintbrush or the actor speaking the role are involved too, and to my mind the only sensible interpretation for AI is that both the prompt-writer and the AI itself have claims of responsibility for what's produced. Nothing else makes sense.

But an AI model, it's argued, can be coaxed into reproducing its training data precisely. This is of no matter because humans can do the same, albeit to a more limited degree. Few of us can reproduce detailed images from scratch (but some people can !), but pretty much everyone can remember a perfectly exact quote or two. And people can create digital-level precision images, for example forgeries. These of course are rightly illegal, but we don't say that because the forger is so good that this invalidates the entire artistic process. 

Going further, of course even copy and paste can be used for theft, but this doesn't mean that this fundamental aspect of the digital world is inherently immoral or invalid. It's the use to which it's put that matters, not the process itself. It doesn't matter if someone has a photographic memory unless they use that to actually produce perfect forgeries which they go on to sell for money.

More often with AI, as with human artists, people don't want perfect recreations of what already exists. They want something new. That AI has a memory inspired by the training data doesn't mean it's stolen anything any more than merely visiting an art gallery constitutes stealing the pictures. Some semblance of the paintings are stored in your brain which, if you're an artist, you can then draw on (consciously and deliberately or otherwise) when you yourself create a new work. Just as with the AI, if you try to create an image in a particular style, you find general trends and apply them to the specifics. Nobody would describe it as theft if, because you once saw a Monet or a Turner in a gallery somewhere, something about your piece had a vague resemblance to a speeding train or some waterlilies.

There is, however, the issue of how the data was acquired in the first place. Exactly how human memory works is still far from understood, but each experience affects us and changes us a little bit, giving us new information from which to draw on. We accept this, encourage it, thrive on it, charging each other money to enjoy our creative works. Nobody says, "you can visit this gallery but you're forbidden to be inspired by anything". But they might well say, "don't take photographs", and of course they would certainly say, "don't take the paintings home with you".

In fact they'd rarely these days have a problem with taking photographs. Taking the pictures off the walls and digitally scanning them without permission would cause greater consternation, if only for the potential damage done. In general the quest to digitise everything is a good one though : it doesn't give you the infinite analogue precision of the original but it can give you enough to be useful. Art which is inherently digital is different, as there the "copy" and the original can be literally identical (caveats about original files often being vector formats notwithstanding). Would this copy then constitute theft ?

My answer is sometimes yes, but not necessarily. If a gallery themselves puts a digital scan online and freely available, or allows people to take pictures, then they have no reason to expect security for the digital copies. They simply have to take it as a given that people might use this for any purpose because to impose specific restrictions on what anyone can access is sheer folly. The obvious solution for controlling access (which I myself use) is to limit the resolution of the images they put online, or paywall it, and forbid taking photographs. And by this token, if AI companies have not paid for the access rights to copyrighted content, then even though what they've done amounts to copying rather than taking a physical product, then it can be constituted as theft. What they've done is not steal the product itself, but steal access rights. This is still taking without permission, i.e., theft*.

* There is of course a very important moral difference between increasing access rights without permission (making a copy, and/or making that copy available more widely) versus taking a physical product – the latter denies access rights to the original owner, which is hardly the same as giving more rights to more people.

It's a bit like paying to access a gallery versus sneaking in after hours for free. Authors and owners surely have some rights to control access, not just the thing itself. Perhaps a better analogy : it's like breaking in to take high resolution scans without permission. But if they're just paid for regular access rights, and/or ingested the vast amounts of freely available online data that anyone can access anyway, and this gets copied as part of the training process... then no, this not theft. Neither the original product itself nor access has been compromised in any way. This would be like erecting a statue in your own front yard and declaring that anyone who took a photograph was a thief, or even saying that if anyone even looked at it then they would be guilty of stealing the image.

That's my take then. AI images constitute art because they express human choices and will, the only novelty here being their particular method. Using publicly available data for the training doesn't constitute theft because it doesn't change the original products or affect their access, provided the AI companies paid for access in the usual way. If they didn't, of course that's theft and they shouldn't have done that.

Seems perfectly straightforward to me.