Statement on AI Art

I would prefer to not be arguing about anything under the All These Roadworks brand, and I would prefer to be limiting my political statements to espousing the rights and safety of women, the importance of positive enthusiastic consent, and the dangers posed to free speech and adult creators by fundamentalist lobbying.

But it’s clear this question is going to come up, from time to time, and rather than have a thousand small arguments in a thousand small fora, I’d prefer to state my position once and move on.

I’ve written a little bit before on AI art. At the time of last writing I had indicated I wasn’t intending to make much use of it. That’s obviously changed a little. The quality of AI art has advanced very swiftly, and I have a better understanding of exactly what it’s capable of and how it works.

My current policy is to use AI art in all contexts where I was previously using copyrighted porn images or Creative Commons Zero images. There’s a phase-out process going on at present, but the intention is that after 28 February, all art used in connection with stories or content created by me will be either AI generated or created on a paid-commission basis.

The short version is this: no matter your beliefs about AI, the use of AI art is unquestionably more ethical than what I was using previously. And I don’t think there’s any reasonable argument to be made about what I’m doing now coming from anyone who didn’t object to what I was doing before. That’s not the whole of the issue, but it’s sufficient to resolve it for most purposes.

I started using porn photos with stories when I first started doing this because everyone does it, and the “caption” culture is a big thing on adult blogging sites. I’m making a full-time income now, and “everyone does it” is not a sufficient excuse, and you’re not going to see those images in connection with my work once the last of the current batch make their way out of the queue.

Creative Commons Zero images provided a better alternative, but not a perfect one, because it’s impossible to satisfy myself that they were validly and deliberately licensed by their original author, or that the humans depicted in such images provided model releases that are compatible with the CC0 licence and compatible with their use in adult/erotic contexts. (Also, practically speaking, it’s hard to find anything genuinely sexy in the CC0 image pool.)

So we come to AI art. And what needs to be said about AI art is that, right now, it is unambiguously legal for me to use the art I am generating in the way I am using it. Whether I own a copyright on that art is a matter of debate in some jurisdictions, but it is unarguably the case that no one other than me owns such a copyright, and there is no legal barrier to me using it, and making money from it. The somewhat customised software I am using is extremely and specifically permissive in that regard.

It’s infinitely better than what I was doing before, and if you didn’t object to the porn photos in the past, the position now can only be seen as an improvement.

So: on those ethical issues of AI art – let’s get into the weeds a little here.

There are two chief complaints about AI art:

ONE: AI models can be used to produce images “in the style of” an artist in a way that is unethical.

“Unethical” is doing a lot of heavy lifting there. Copyright laws are very clear that you cannot copyright a process or a style You can only copyright the results of that process or style. And the law doesn’t take that form because of some oversight or accident. Those laws are some of the most analysed and heavily argued laws in the world, and they’ve ended up in that position because artists wanted it that way. (I don’t pretend that ALL copyright law is what artists wanted, although it certainly is all what corporations wanted.)

This is actually something that corporations would love to see changed. Disney would love to copyright “the Disney style” and “the Pixar style” and stop people ever making an animated film that looks remotely similar to the Disney output. Adobe would have loved to copyright many of the processes that Photoshop enables, and have an exclusive licence to create art via “layers” and “filters”. If Andy Warhol could have copyrighted pop art, he would have, to make a statement, it nothing else, and it would be a style forever denied to anyone who didn’t pay money to his estate. Anime artists would be forever worried about receiving a suit for infringing on “the Ghibli style” or “the Gainax style” or “the Clamp style”, and that’s assuming that Oten Shimokawa or one of his peers hadn’t copyrighted “anime” in general as a protected style.

We do not want to forbid generating images in the style of another artist. No one wants that, and least of all artists.

But I, personally, am not generating art in another artist’s style anyway, so this limb doesn’t apply to me in any case.

Let’s turn to the other complaint.

TWO: AI models are trained on images, and the authors and subjects of those images did not give permission for those images to be used in that way.

Right now, intentionally or not, that’s legal. Of course it’s legal. We train humans in exactly this way, by taking them to art galleries, by showing them good art and bad art, by getting them to take in as many images as possible and draw conclusions from those images as to how to communicate in the visual medium. We pay no licence to those artists to learn from their images.

A person can take those learnings and use them to construct an algorithm that generates images in certain ways. Why should it be different to have a machine construct those algorithms through a largely identical process? The end process is an algorithm – a kind of subjective “truth” about how to produce images – and the model contains only that algorithm, not the images it used to derive it. It knows what “a painting by Rembrandt” looks like. So does Google. So do I. It expresses its knowledge visually.

But okay, let us follow the thought to its conclusion. Say that developers are forbidden to train AI on images without purchasing a licence to those images. How does that play out? Does it end AI art? Does it place money in the hands of artists?

No. The demand and usefulness of AI art is simply too overwhelming. The money is there, and sufficiently motivated people will be willing to pay it. But that money will be paid to corporations who possess huge image libraries, such as Disney, or Paramount, or Amazon, or Google, or Meta, because ain’t nobody got time to go make individual deals with a million small rights-holders. None of that money will flow back to artists.

And the increase to the cost of generating the models will mean that instead of being developed in the public space, and freely made available to the public, these models will be developed and held by corporate interests, to be used in corporate ways.

By all common sense, small artists SHOULD be the greatest beneficiaries of AI art. The most effective ways of using it are by pairing it with existing digital art skills to modify and customise the output that the model produces. Artists can produce more and better art using AI than anyone else, providing they put the time into acquiring and understanding the tools. But raising the cost of training AI models will raise the cost of gaining access to those tools. Artists would be shutting themselves out of their own market, and largely yielding it directly to corporate interests.

Or do we ban AI art altogether? Some communities are trying to do that, and immediately running into two key problems. One: it is simply not possible to reliably recognise AI art. Much high-quality digital art looks like AI art, when it is not. Much high-quality AI art looks like traditional digital art. That gap will only narrow with time.

And two: digital artists have already been using AI for years. A great many filters, brushes, upscaling tools, and other digital image applications are built on AI, using predictive techniques to understand what an image should look like, often by reference to other images. There is no simple line between “AI” and “not AI” in digital art, no easy place where the objectionable content begins, and that’s before taking into account things like running an AI model to generate inspiration for a piece that is executed entirely with traditional methods. We have already embraced the basic concept of computer-assisted art, and there is no way to disentangle that in a way that’s sufficiently clear to base enforceable laws around.

With all sympathy to small artists, AI art exists, and there is not any regulatory or policy position in relation to it that will place small artists in a better position than they currently are.

I understand the emotions. It is shocking, and scary, and it WILL replace a good portion of the transactional art market, in much the same way that industrialisation did to the pre-industrial artisans, or the way that the printing press did to scribes, or the way the photographic camera did to portrait painters.

I have personally worked in an industry that was rendered largely obsolete in my lifetime by digital automation. I saw it coming and I prepared for it, identifying the skills that would remain valuable and the ways that I could personally adapt to the future. That’s all you can do.

Don’t get me wrong: this kind of automation is NOT a simple “good thing”. It’s disruptive, and people get hurt, who do not deserve to be hurt. We may end up better off as a world, but that doesn’t diminish or excuse the real human cost of that transition. But it *is* going to happen. The technology is in hundreds of thousands of hands already. It doesn’t just stop existing.

What we can learn from history about these kind of disruptions, though, is that the amount of good we get from them is directly proportional to how well we keep the power in the hands of average citizens rather than allowing it to be concentrated in corporate hands.

In looking at AI art, we MUST take a path that avoids raising the transactional cost of developing and using the models. We MUST take a path where the models can be freely and cheaply installed and operated by anyone. We MUST take a path where any artist, from a child to a professional, can access these tools, and learn to use them, and own the art that they create with them.

The absolute dystopia outcome is this: that only corporations can build and run these models; that they dominate the transactional art market through quality and efficiency and completely displace small artists from that market; and that even when small artists make use of the AI art tools, the corporation who creates or owns the model owns the output they generate from that model.

Being scared of the change will not stop the change. AI art is going to be a huge part of how humanity creates visual images into the future. The choice we have ahead of ourselves is only who benefits from that change – and in that context, you need to be aware that the more the space is regulated and legislated, the less likely it is that the ultimate beneficiaries will be anyone other than corporations.

This is my definitive statement on the issue… for now. It’s almost certain that someone, somewhere, will legislate, and legislate badly, and that’s going to suck for everyone. It’s also likely that the technology will continue to evolve. Eventually, it becomes almost certain that large corporate interests will take the field for one reason or another, with a killer-app product, or to advocate for some law or other, and that will be significant. All I can say is that when things change, I’ll change with them, and speak again.

• All These Roadworks
• 13 January 2022