The PIG Menace

Two tiny figures fishing in a small boat on a lake in an alien landscape. A domed, science-fiction tower rises in the background on the left of the scene. To the right, before a sky full of stars and planets, a huge spaceship in the shape of a grey alien skull looms menacingly over the pair.

I find the term “AI art” offensive because it’s not only a provocation, but also inaccurate in every particular.

Art involves, in the first place, having something to express, and then a creative process where decisions are made about how best to express that thing. Talent and skill are optional – you don’t have to be good to create art, and that may be one of the most inspiring things about it.

A scene of a freeway coloured in loosely scribbled felt pen on yellow paper. Beyond the road a grassy bank rises, topped with red buildings, including a McDonalds. The freeway bulges toward us, giving the scene a dual vanishing point perspective. Two buses are incredibly detailed, one with a cigarette ad reading 'NOW IS LOWEST' on the side. Other cars are abstracted to the point of being little more than extruded trapezoids. There are big white clouds in the sky. On the bank a dedication identifies the artist, location and the date as 1994. Wesley Willis wasn’t ‘good’; merely amazing. Maybe PIG could copy his style, but nobody else, man or machine, could come up with it.

As for “AI” – a term which adds mystique as well as menace to what is ultimately just an incremental bit of automation – what’s being used here isn’t any kind of intelligence but a combination of machine learning of visual patterns, large language models, and procedural generation. What we’re seeing is just the evolution of Photoshop’s “content aware fill” or autocorrect on mobile phones.

So let’s call it what it is: Procedural image generation. PIG.

Pull its tail and the PIG will cough up a picture for you.

Dangers of Machine Learning

Sandra has written a good summary of the arguments against the current crop of ML. The most serious don’t relate to art – its use for surveillance, including deadly surveillance such as the ‘Gospel’ system currently being used in the Gaza genocide; the impact of training new ML models – an energy-intensive, brute-force process – on climate change; and the concentration of control over the technology in a few hands.

The art-related objections include the idea that it calls into question the meaning of human art. The more I interact with it the more I feel this isn’t an issue; it isn’t capable of any kind of meaningful artistic process, and this omission isn’t simply a weakness that can be trained out of it, like teaching it to give hands five fingers most of the time; it’s inherent to the way it works.

It fills space up with colour and texture according to rules: A goes next to B, the prompt asked for X and not Y, etc. The results can seem like magic but that’s simply because there are a lot of rules in these models. It has no memory of its own; it has to simulate memory by looking at part of the log of interactions every time it processes a new request. When asked about something that falls outside its context window, it just makes it up anew.

It’s a topsy-turvy technology. That is to say, it upends our expectations about both computer behaviour and human skill. We expect computers to be linear, reliable, good at facts and copying and stable, repeatable behaviour. We don’t expect something that won’t give the same answer twice, hallucinates constantly, operates more on the level of vibes than facts, and routinely violates the rules it’s been instructed to follow.

This topsy-turvy quality extends to the quality of the illustration it produces. It’s amazingly good at the hard stuff – the folds of clothing, light and shading, producing something that looks incredibly sleek and polished – things that in a human would be signs of long years of practice. It’s terrible at some very basic things – if I said “draw me six people, two laying down, two facing away, one in profile, one facing the camera and wearing a blindfold” a non-artist human would do a better job than the PIG.

A human artist has a range that the PIG can’t match. It has to follow paths laid down by real artists, and it can move around adjacent to them and between them but it can’t strike out on its own, because its entire schtick is to follow rules it’s learned by observation.

Someone told me that artists would be doomed if they didn’t start training PIGs on their style and having the machine do the drawing right now, because artists who can work at machine speed will outcompete the ones who can’t. But this hasn’t happened and I don’t think it’s going to, because of the belt-buckle problem.

If you have a unique style, and you get a commission which requires drawing a character wearing a belt, you can either draw dozens of belt buckles teaching the PIG to draw a range of buckles in your style, or you can just take a few seconds to draw the buckle yourself.

A real artist is faster, both because they’re already trained, and because they can just, metaphorically speaking, make a cup of tea, instead of generating 600 cups of something which is almost, but not quite, entirely unlike tea, and then having to sort through them and figure out which one is the least unlike tea.

Which brings us to the next art-related issue, that of plagiarism. It’s a common criticism that AI is just copying other people’s work, and one can point to the inclusion of (garbled) versions of artists’ signatures on generated images.

But this is rather evidence for the opposite – we expect computers to copy, but PIG is very bad at copying because it can’t remember any of the art it’s seen. All it remembers is the rules it derived from observing that art, including the rule that art of a particular style often seems to have a little squiggle in the corner. This shows that it’s imitating, not duplicating.

But the imitation itself is objectionable to artists; and it's perfectly understandable that you don't want a horrid little robot run by a horrid little techbro biting your style. There’s a genuine dilution that arises from every idiot who types in “cool skeleton mike mignola style” as a prompt polluting the image search results for people looking for actual Mignola art.

There have been calls for some kind of legal protection from having art scraped and used to train PIG models, but the biggest IP holders are highly litigious corporations who are very good at getting laws drafted to advantage them. If an artist can sue an ML company for having a machine observe facts about their art for training, what happens when Disney can wield the same power against people who, in their opinion, have learned something by observing Disney’s work-for-hire IP?

If something can imitate you, there’s the implication it can replace you – a fear which hasn’t been allayed by the fact that when artists complain about PIG they are, as my partner put it, surrounded by horrid little men with bored ape avatars, dancing in a circle and singing joyfully about how the artist is going to lose their job and starve in a ditch and they deserve it for being a luddite.

This isn’t an unreasonable fear. If corporations love two things, it’s cutting payroll and attracting venture capital by jumping on the latest idiotic hype bandwagon, even when that bandwagon is blockchain, a technology with more-or-less zero useful applications.

PIG & FRP

PIG can be a genuinely joyful technology; the ability to conjure up an image that looks kind of like what you imagined in seconds is a wonderful magic trick. But we need to maintain a realistic assessment of what it can and can’t do, and develop an ettiquette and an ethics in how we interact with it.

D&D campaigns don’t have an art budget. Maybe they should – my Saturday game collects £20-40 each week to pay for the venue, which over 2 years of play could add up to £1500 or so – so a group playing at home could set aside a similar amount to pay an artist for character sketches, tokens, illustrations of dramatic moments and the like.

It’s also a delight when there’s an artist at the table and you can look over and see what they’re doodling during the game, a lot of the campaign ends up getting illustrated organically.

Personally, I simply would not use procedural images in a commercial product. Not only because I’d rather work with an artist in the first place, but also because I believe people are savvy enough that they’ll quickly build up an association between PIG illustration, shoddiness and corner-cutting. It may look very slick and polished now, but people are good at building up associations – it’s how our ancestors learned which fruit not to eat – and that very slickness and vague genericness of generated images will quickly come to work against those using them commercially. It doesn’t entirely matter what humans can do that ML can’t – whatever it is, people are going to gravitate toward it, because the people making the heaviest use of PIG are going to be chancers like those behind Willy’s Chocolate Experience.

But for anything where you’d otherwise be going to grab a picture off google images? Anything that’s for free or for fun? Illustrating blog posts, minor NPCs from your game, going back and forth helping to firm up your idea of your next character – these are all places where it can enrich the FRP experience.

So I’m enjoying playing with it, but these are the guidelines I’d follow when using it for FRP illustration: