By Jo Lindsay Walton
NightCafe AI’s response to “Sunflowers, Van Gogh”
Klara and The Sunflowers
This issue’s cover was created by an AI. Or … was it?
Machines have made art for a long time. In the mid 19th century, John Clark’s Eureka machine was dropping perfectly okay Latin hexameter bars on the daily. Harold Cohen’s AARON began scribbling in the 1970s and sketching plants and people in the 1980s.
But with the likes of MidJourney, DALL-E, Stable Diffusion, Disco Diffusion, Imagen, and Dream by Wombo, 2022 marks the start of a new era. These AIs accept natural language prompts and produce often startling images. Suddenly the conversation has shifted from what little the AIs can do to what little they can’t do.
The AIs can’t paint complex scenes with many parts, for instance. You’re better off generating the pieces separately and then jiggling them together in Photoshop or GIMP. They can’t paint eyes terribly well, unless your subject happens to be a stoner ghoul. If the moon shines behind your subject’s head, it often bulges strangely, bearing ominous tidings for tonight’s high tide.
Still, the AIs are getting better all the time. Some online art forums are already inundated with spam. There have been instances of AI users setting themselves up as freelance artists, claiming to create the images themselves using traditional methods (Photoshop is now ‘traditional methods’! We are definitely in the future).
Worse still, the rise of AI art has led to the rise of the AI Art Bro. These combat philosophers, who perhaps recently cut their teeth extolling NFTs, love nothing more than to troll freelance artists nervous about next month’s rent. Yet it would be unfair to write off AI art just because it has some disagreeable advocates. Luckily, as science fiction writers and fans, we’re well-equipped to make more nuanced assessments.
Or … are we?
The uncomfortable fact is that science fiction hasn’t been amazingly good at illuminating the ongoing AI revolution. With notable exceptions, we focus on questions like, ‘Can an AI think? Feel? Love? Dream? What does the way we treat machines tell us about how we treat one another?’ These are enchanting and perhaps important questions. But they tend to overshadow AI as it exists within data science and critical data studies, and the huge role it is already playing in everyday life. So maybe science fiction writers could do more to infuse our work with an appreciation of AI as it actually exists?
Putting the art in artificial
Text-to-image AIs, like most of the AIs making waves these days, are examples of Machine Learning. Machine Learning relies on immense processing power and lots of pre-existing examples to train the computer through various kinds of complex trial-and-error processes. Machine Learning is sometimes described as automating automation itself.
The truth is, nobody has the faintest idea how to write a computer program that will paint any image you ask it for. But we do know how to program a computer to program itself to do so, by carefully guzzling up millions of images created by humans.
Could that, in part, explain why the moon grows so bulbous as it passes behind a portrait subject’s head? Are the lunar contours haunted by the flowing lines of capes and crowns, helmets and halos and hairdos, that attend so many of the heads in their training data? Must moonrock budge, just a bit, to accommodate the tumbling locks of the windblown studio supermodels?
I’m not sure. One thing is clear. When you take into account all that training data, the contemporary debates about AI art start to look very skewed. Pundits ponder whether these AIs are mere tools or artificial persons, and who should own the artworks: the user of the AI, the owners of the AI, or perhaps even the AI itself? But what about all the human models who taught the AI models a trick or two? And what about the millions of human artists, illustrators, photographers, and designers who also contributed the training data?
Legally speaking, artists who have contributed to the training data appear to have very limited rights. Of course, law is created through particular judgments, as much as it is through legislation. The current wave of AI art hasn’t been seriously tested in court, so the answer to any complicated question about how copyright law applies is, “We’re not sure yet.” Nonetheless, the EU Digital Single Market Directive has created a copyright exception which means that it’s okay to use images, without needing their owners’ permissions, for training an AI model. The UK has something similar in place for research purposes, and looks set to extend it for commercial purposes too.
Despite all this, many artists are enthusiastically taking up these enchanted paintbrushes. Pros before bros: those skilled in the arts of Art can wield AIs to far greater effect that ordinary users. Yet at the same time, many are pointing out the double threat the AIs pose. Artists may well have to put up with their artworks being fed into models, so that these can imitate their styles and subject matters. Artists may also end up competing with those same AIs for clients and commissions.
In short, support artists! For those who have cause to work with artists—for your book covers, for your websites, for your D&D or Lancer campaigns, for your kitchen wall or your next tattoo—the coming months and years are a
great time to show a bit of extra solidarity, and to funnel some cash into a field full of upheaval.
‘Here is some information I found about “Pod Bay Doors,” Dave.’
Compared with visual art, AI text generation has made a smaller splash. But its ripples carry at least as far. What do you think, BSFA? How would you feel about an AI-drian Tchaikovsky or an Una McAutomatic, an Ursula K. Le ImageGen or a Samuel R. DALL-E? Adam Roberts wrote a collection called Adam Robots; what if Adam Robots wrote a collection called Adam Roberts?
Then again, readers of the future may not know when an AI has been involved. One gloomy vision is this: in the future, an ‘author’ could be anyone with enough wealth and pomposity to employ a small army of AI trainers, AI prompt engineers, market researchers, focus group moderators, beta readers, sensitivity readers, narrative troubleshooters, specialist copyeditors, astroturf fans, and so on, to implement their own Deep Literary Genius. Not to mention a few lawyers to bully other authors away from whatever they consider their turf. They might do much of this discreetly, even secretly.
That’s not a prediction. It’s just one possible timeline, and there are much nicer timelines we could steer towards. Like what? My hunch is that the best futures, when it comes to the vocation of writing, are those where copyright law is radically transformed.
Some writers can be pretty conservative when it comes to copyright. Copyright makes money for some of us, and a world without it is hard to imagine. But many writers at least recognise that copyright law doesn’t really reflect the messy realities of our creative processes. A host of legal concepts — originality, ideas and expressions, scènes à faire, labour and skill or judgement, fair dealing or fair use, joint authorship, among others — are no more than flexible approximations of those messy realities. Writing just isn’t like what copyright law pretends it is like.
Then again, the legitimacy of copyright law doesn’t really rest on any claim to detailed accuracy. Copyright law only claims to approximate the practices of creators well enough to foster an environment where creative experimentation is rewarded, and one where the fruits of creativity are shared fairly between the individual creator and wider society.
But the more that our creative processes are transformed by AI, the more they diverge from how copyright law imagines them. This means that copyright’s claim to legitimacy grows more and more tenuous. The whole
rickety, jury-rigged apparatus grows more visible. We see its flaws, and not just in relation to these emerging technologies—we see how copyright law has been flawed all along.
Today’s AI images would not be possible without the countless humans who contributed to the training data. This also serves as a reminder that every work of art is nourished by countless humans. From editors and agents and writing group comrades, to the authors of the books that have filled you with fire and ice, to the folks who recommended to you or bought or lent you those books, to the amazing weirdoes who inspire your characters, to that specific wonderful quip you transferred from your witty best friend’s lips to the charismatic sneer of your space corsair, or that cup of tea your dad set down on your desk when you needed it most … these others may not all be the author of your work, but nor are they nothing to it. The idea of trying to keep track of all these little tributaries feels silly. Yet those whose contributions have not been registered in any way have disproportionately been women, people of colour, and working class people.
There is a lot of talk about futures these days. ‘Urban futures,’ ‘energy futures,’ ‘climate futures,’ ‘transport futures,’ ‘work futures’: the idea is, we can actively shake off our presuppositions about these things, and envision futures worth fighting for. In that spirit, can we also imagine better ‘writing futures’?
Science fiction writers and fans may be well placed to do just that. What might it take for human writers to flourish side-by-side with sophisticated text-generating AI? Would it help if we could publish stories about any fictional characters we want, Wanda Maximoff wandering the mean streets of Ankh-Morpork, Jean-Luc Picard leading an away team to Annares? Could there be upsides, under certain circumstances, to other creators playing with the characters you invented? Or humans and AIs collaborating to share characters, worlds, stories, ideas, plots, styles? Do we want a free-for-all, or some sort of radically different copyright regime? How might literature and arts be funded more generously and fairly? What alternative compensation mechanisms to copyright exist, have been proposed, or are newly imaginable? What might it truly mean to give credit where credit is due?
Through all these questions runs a deeper question: what gets us closer to a world where anybody who wants to make stories or art can do so, and nobody has to worry about selling those stories or that art, or wowing funding gatekeepers, just to put food on our tables and roofs over our heads? And what pushes us further away from that world?
And I wonder if we could go even further? Stewartificial Intelligence Hotston lays out in his guest editorial this issue’s theme of justice, and makes an urgent appeal for new ideas. And as Gautam Bhat-AI suggests in his article, science fiction might play a role in exploring distributional justice. So I wonder, might this be a moment where not just the nature of Intellectual Property, but the nature of all property, can be subjected to fascinating and ferocious reimaginings?
BSFA Committee News
BSFA members will have noticed some comings and going announced in the monthly newsletter. We say farewell to Writing Groups Officer Terry Jackman and Councillor Yen Ooi, with huge thanks for all their hard work and general brilliance. We are delighted to welcome Stewart Hotston as BSFA Councillor, (Stewart also guest-edits this issue of Vector), as well as Writing Groups Officers Mark Bilsborough and Sam Fleming to fill Terry’s shoes (one shoe each, they claim).
As we go to press, we also hope soon to be announcing a new Awards Officer or two. Following a motion at the 2022 AGM, a working group has been set up to explore potential changes to the BSFA Awards — something will be presented at the next AGM, likely including at least one new category. If you have thoughts you’d like to share about the BSFA Awards, now is an especially great time to share them (Farah Mendlesohn at email@example.com is the best person to direct these to).
If you don’t already, don’t forget to follow the BSFA Twitter account (@bsfa). There’s also a Facebook group, a Discord server, the main BSFA website and the Vector website. We’d also love to hear (firstname.lastname@example.org and/or email@example.com) your thoughts on the BSFA’s digital presence.
 Yes, it definitely was! It was also created by humans though, including Polina Levontin, the human who used the AI, the humans who programmed and trained it the AI, and the humans who created the training data.
 NFTs, Non-Fungible Tokens, a form of blockchain finance which is entangled in complicated ways with the art world. Currently most famous example is the Bored Ape collection. NFTs are a whole other conversation.
 See for example Ruha Benjamin, Race After Technology: Abolitionist Tools for the New Jim Code; Kate Crawford, Atlas of AI;Meredith Broussard, Artificial Unintelligence: How Computers Misunderstand the World;Safiya Noble, Algorithms of Oppression: How Search Engines Reinforce Racism; Simone Browne, Dark Matters: On the Surveillance of Blackness; Nick Couldry and Ulises A. Mejias, The Costs of Connection: How data is colonizing human life and appropriating it for capitalism; Shoshana Zuboff, The Age of Surveillance Capitalism. But no doubt these generalisations are unfair. What great science fiction have you encountered (or written!) about Machine Learning and data science?
 Machine Learning is a very broad term. The leap in image generation we’re seeing with MidJourney etc. is to do with a shift from GAN models to diffusion models. The diffusion model trains itself by taking an image and adding more and more random noise to it, remembering all the steps along the way. Once it has an image of ‘pure’ random noise, it can practise denoising the image, checking its progress against the sequence it has saved. It does this many many times — say about 150 hours and $600,000 of compute cost — and each time adjusts the weights of its neural network. What you end up with is a neural network that is pretty good at ‘finding’ any given image in a random field of noise, even if the image wasn’t really there to begin with. Of course there is plenty more to it than that, but hopefully it’s forgivable as an intro. In the same vein, there’s a xkcd comic that depicts Machine Learning as a pile of linear algebra with a funnel to pour in data, and a box that spits out answers. ‘What if the answers are wrong?’ ‘Just stir the pile until they start looking right.’ xkcd.com/1838/
 Should people whose likenesses appear in the training data have any rights? Here the question is partly about identifiable individuals, especially celebrities and public figures, whose appearances are becoming available for use and manipulation in new ways. But it’s also about the visual representation of groups — it’s about race, gender, nationality, class, and other big categories, and how these get inscribed into AI models in ways that may reinforce or (perhaps) subvert stereotypes and biases. Or maybe even give rise to new ones.
 One area of controversy is whether AI art should be copyrightable. Many jurisdictions say that it definitely is not copyrightable, whereas the UK’s Copyright, Designs and Patents Act 1988 leaves things much more open. It’s worth emphasising three things. (a) Creating something non-copyrightable doesn’t mean you can’t sell it (it just means you can’t stop others from selling it too). (b) Combining non-copyrightable elements can give rise to new copyrightable works. (c) It may be tricky to prove what has or hasn’t been AI-generated. So you might get situations where somebody protects an AI-generated work using copyright, simply by insisting it was actually human-created; conversely, you might get predatory appropriation of human labour under false claims of AI authorship.
 There is the opportunity to opt out. It is easy to see what this might mean in the case of a training dataset (just remove the images), less easy to see what it would mean in the case of a model that has already been trained on that dataset.
 My first real experience of using AI art was creating the Wildlaw Judgement Generator, a small artistic collaboration with legal scholar Bonnie Holligan and artist Elias Youssef, which also included surreal courtroom sketches generated using Dall-E Mini. I feel pretty good about how it all fit together: a big reason it happened in the first place was getting some funding to pay Elias; the AI was used in ways that felt like it set off the human art, rather than displacing it or drowning it out; the glitchiness of the AI didn’t feel like an irritation to be eradicated, but more like a source of inspiration and fruitful weirdness. The Wildlaw Judgement Generator also relates to this issue’s loose theme of justice, specifically environmental justice. It is inspired by an actual legal case, which explored how much (if any) compensation should be paid to a commercial fishery by the Environmental Agency, after the Environmental Agency placed a heavy restriction on how many fish could be caught, in order to protect the fragile fish stocks. The bigger question is, as climate change and environmental crises make many previously viable activities no longer possible, how should the costs of transition be distributed? It’s also part of a bigger project (the UK Earth Law Judgments project), reimagining legal judgments from an ecocentric perspective. https://sadpress.itch.io/wildlaw-generator
 For starters, even the existing regime of copyright law unsettles some common convictions about how property works. For instance, it is possible to create something that is both original (so you get copyright) and yet infringes on somebody else’s copyright (so you can’t use your copyright). Some creations (like the sentence, “How are you?”) would be so absurd to privately enclose, they automatically belong to everybody. Then there is the law on fair dealing, which says it is perfectly fine for somebody else to use something that is ‘yours’ so long as it is for one of a specified list of purposes in the public interest. What if there were a few more fair dealing exemptions in the physical world, on the ownership of land and factories and finance?
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License