Over on Facebook the other day, Tony Dowler, an artist whose work I admire (particularly his maps), posted this:

Yesterday, I asked AI to produce some cover art for my games. The results were pretty astounding. I immediately thought that these were probably better covers than I could have created myself. It left me with a huge dilemma.

I come across art that’s vastly better than mine every day. When I see an artist who exceeds my capability, I strive to learn what I can and then to make art that’s entirely different. I don’t want to aim for a goal that I will always fall short of, but I also want to get everything I can to improve my own art.

I’ve got a response for when other artists are better than me. Every artist has to have that. It’s a survival mechanism. But when the better artist is an AI, I don’t know how to respond. I don’t whose technique I’m learning from, which makes it feel like I really am stealing. I can’t take comfort in the fact that some other artist can make the thing I can never make, because it’s not another artist doing it.

Is it time to go back to analog art, and making things that are as messy and analog as I can make them? Should I accept that AI is my source of reference art and inspiration now and that human artists are out of the loop?

I've heard at least one of these points raised by artists over the last few months, and I asked Tony if I could use his post as a jumping-off point to talk about the miasma of issues around art, artists, and AI imagery.

AI Supremacy (aka The Oh Shit Moment)

I'll address this point first and most briefly because I've covered it before. In the intervening time since I researched what AI was capable of in terms of "creativity" - very little has changed. If anything, I'm less impressed now than I was a year ago with the capabilities of generative AI.

While the resolution of AI generated images has gone up, the range of images that AI is capable of generating seems to have fallen into very specific niches: concept art renderings, stock photography, product photography (for things that don't exist), and Pixar-esque characters. I'm not going to go in-depth into any of these buckets, but suffice it to say that if you write a prompt that touches on any of these areas, a generative AI like Midjourney will likely knock your socks off.

As I've said before, however, a large amount of the wow factor of generative AI is in our very human ability to bend our expectations to what's generated instead of the generated images reaching our expectations. Go through the gallery page of any generative AI engine, and while you may be flabbergasted by what's been generated, if you take some time to read the actual prompts, you realize that almost none of them meet the original request.

Here's an example I selected pretty much at random:

I mean, super cute, right? If I was an artist and asked an AI to generate a cartoon fox, and I got this, it'd hit pretty hard. But let's look for a moment at the actual prompt:

A red fox sitting in a a grassy field, surrounded by a vast expanse of greenery and flowers. The sky is blue. The fox is wearing a monocle with a chain hanging from it on her right eye. Her fur is gently swaying in the breeze, and the sun is shining down on her, casting a warm glow, delicate skin, beautiful hair, large eyes, three dimensional effect, enhanced beauty, feeling like Albert Anker, feeling like Kyoto Animation

My first real issue with any generative AI is the ability to input an artist's name. For what it's worth, I think artists' names should be off-limits. But this image has no monocle on a chain, which seems like it's the main thing this person was going for, apart from delicate skin – which, to be honest, on a fox, I'm not sure what that looks like.

But let's look at something else:

This is the first page of the Google image search for Albert Anker. Our image here pops up on the first page at the 'AI Art Shop,' but if you look at the other images (which actually are Anker's work), you notice that Albert's style is nothing like that image.

I try to get this across to artists who feel blindsided by generative AI: It can't do a whole hell of a lot, and even the things it can do can't be done consistently. Oftentimes, when someone is taken aback by what AI has generated, there's a healthy dollop of 'good enough' thrown in, or they're working with a blank canvas idea, and whatever comes up is acceptable because there are no real criteria. Give generative AI any criteria or any list of requirements, and it falls short consistently.

What Can Artists Learn from Generative AI?

This is a much more personal question and one that I don't feel qualified to discuss much because while I dabble in art and design, I wouldn't label myself An Artist. Having said that, I've spent a fair amount of time delving into the generative AI abyss, and I can offer a few observations on what I see as AI's weak spots.

Cultivate a Distinctive Visual Style

Part of what made Tony's message interesting to me is that I consider his work very distinctive. It's very organic, with excellent ink work and a real sense of Tony's hand (his very real, physical hand) creating it.

While it is true that artist names have been rolled into AI models (which I think needs to be blocked from being accepted in prompts), as we've seen, that style is rarely accurate, and the more distinctive the style, the less accurate generative AI is in the final product.

Move Away from 'Stock' Images

This applies more to photographers than other visual artists, but moving away from stock photography and spot art is a solid creative move. Generative AI is extremely good at images with an isolated, central subject since it requires the least nuance to incorporate into an overall image. It's not as good with interactions between subjects, complex backgrounds, or complex subjects. I know there's always been a market for the character portrait artist - but generative AI can eat that market for breakfast.

Again, not being a serious visual artist myself, my understanding of the challenges that generative AI throws at you is second-hand. Like any technology, though, it has its limitations and increasingly can be spotted by the general public for what it is. Understanding those limitations and exploiting, for lack of a better word, your messy humanity will keep you ahead of the rising tide.

I was already in the process of writing a response to this Big Technology article by Alex Kantrowitz when news of Sam Altman's ouster from OpenAI was announced Friday, and it took me until today to reorganize myself enough to respond. That was enough time to find out that Meta had abandoned their AI safety project and for Sam Altman to be spared the ignominy of having to file for unemployment by joining Microsoft.

Altman's departure for not being "consistently candid in his communications" with the board (whatever the hell that means) is not essential. That the helm of one of the largest (if not THE largest) players in the AI space is suddenly a chaotic mess is critically important. There have already been mass resignations in the wake of Altman's departure, most notably co-founder Greg Brockman, who Microsoft also snapped up after a whole weekend of considering his options. Any sudden change at a company worth more than the GDP of more than half the world's nations is worth at least raising an eyebrow about. Add to that the care and feeding of the most widely used generative AI ... and you might want to raise both eyebrows.

We'll wait and see what all the fuss is about. Still, I'm willing to put my money down on the board being concerned with Altman's messianic view of what OpenAI was engaged in, as revealed in a Vanity Fair article that came out November 15th, but that's just a personal theory.

[Update: As of the day this article was published, Sam Altman was back at OpenAI after literally at the company turned in their resignations. It appears my theory was at least partially correct: Sam Altman is capital-first with AI and was ousted poorly by decelerationists on the board. They have since all been shown the door. I therefore anticipate OpenAI to push the accelerator to the floor.]

Meta quietly dismantling their 'Responsible AI' program is just par for the current accelerationist course. With billions flooding the zone for anything and everything AI, every dollar not spent directly on training and expanding LLaMa (or proprietary AI built on the backbone of LLaMa) by Meta is, of course, 'wasted.' Right now, AI is all-gas-no-brake, which seems unlikely to change anytime soon, as the slip-n-slide of capitalism sends us to whatever bottom Silicon Valley can conjure up.

Which brings me to the article that first spurred me into writing, AI Doomers Are Finally Getting Some Long Overdue Blowback:

Worrying about AI safety isn’t wrongheaded, but these Doomers’ path to prominence has insiders raising eyebrows. They may have come to their conclusions in good faith, but companies with plenty to gain by amplifying Doomer worries have been instrumental in elevating them. Leaders from OpenAI, Google DeepMind, and Anthropic, for instance, signed a statement putting AI extinction risk on the same plane as nuclear war and pandemics. Perhaps they’re not consciously attempting to block competition, but they can’t be that upset it might be a byproduct.

This paragraph manages to touch just about every argument that accelerationists leverage against the notion that we are running a sprint (or a quarter mile of asphalt) instead of a marathon when it comes to the development of generative AI:

• "worrying about AI safety isn't wrongheaded" - but it is, at best, secondary, tertiary, or quaternary. We all know which priority is Number One.
• "They may have come to their conclusions in good faith" - of course, the sobriquet 'Doomers' lets you know how seriously they take those good-faith conclusions.
• "companies with plenty to gain [...] have been instrumental in elevating them" - and here we have the argument that is the crux of the article, and the crux of this argument: No matter how well-intentioned, any regulation of AI is just handing the future to OpenAI, Meta, and Google.

Let me be exceptionally clear: I am a skeptic of AI, and a decelerationist only at my most vehement. I don't harbor dark nightmares of Skynet, Gray Goo, or Colossus when I consider the dangers of the future of AI. I hope most AI decels share the same rational level of concern. There is a difference between "we're all going to die tomorrow if Artificial General Intelligence comes online" and "generative and general AI have social and economic impacts that we, as a capitalist society, are both ill-suited and unlikely to address."

For all of the hue and cry about the dangers of regulatory capture brought on by decelerationist concerns, we already know that regulatory capture is highly unlikely; why? We already have a perfect example of a technology with significant social and economic ramifications: social media.

At the time of publication, we're 19 years into a giant social and psychological experiment run by a handful of companies shaping society and possibly human cognition in ways we have yet to understand, let alone get substantial control over. Some of those companies are owned by terrible people with agendas I feel safe in labeling asocial at best, if not outright hostile to the project of liberal democracy. We've seen mountains of evidence that even without direct experimentation on the part of these companies (which we know they have ALSO done), their industry is ripe for regulation, if not investigation - and we have not had a single meaningful piece of legislation in place for nearly a decade and a half.

This lack of regulation hasn't resulted in a proliferation of social media platforms and robust competition - in fact, given the state of play in social media, you could have sworn that the wrath of Regulatory Capture had been invoked in the social media sphere. The only growth area in social media networks is right-wing grifters peeling off silos of folks who can't live another minute without using the N-word in public again. I'd put a stake in 2011 as the last major launch of a 'new' form factor of social media (live streaming via Twitch, since acquired by Amazon) that has given us any meaningful innovation in the space. So, despite zero regulation by any governmental body, we've seen nothing but "Facebook for/with <INSERT NICHE HERE>" or "Twitter except <INSERT FEATURE HERE>" for the last decade at least. So don't tell me that regulation will stifle innovation and lock in some oligarchical high-rollers. Silicon Valley has managed to lock in a handful of social media giants all by itself, thank you very much. Even the latest contender, TikTok, is just a social media giant from another country that's managed to get a foothold here. (TikTok is also just Longform Vine. Rest in Power, king.)

Social media also gives us a template by which we can gauge the potential downsides of AI as a technology. You might be able to find "thought leaders" who will still lean into the idea that social media has been a net positive for society – but even the most fervent booster of social media would have to admit that positivity is marginal. Since 2016, we've known that social media (and the tools and analytics that allow micro-targeting online populations) is, in the wrong hands, capable of shaping events to the self-interested few instead of the whole. The pandemic turned the volume on misinformation channels up to 11, and we've not seen it abate since.

Now consider the following points:

1. The same coterie of companies that have served as conduits of misinformation, knowingly or unknowingly, are the same companies currently invested in pursuing generative and general AI.
2. We have not seen any meaningful attempts at correcting the ample problems that already exist with a technology that will be 20 years old in 2024 and even outright resistance to harm-reduction on the part of social media (I'm looking at you, Elon).
3. Generative AI offers a far more robust suite of tools to the self-interested few than social media. These are tools capable of not just spreading misinformation but also generating it wholesale, at quantities that defy the ability of any concerned party to mitigate the damage caused.

The accelerationists propose that the same actors who have profited off the most toxic corners of social media are those we should put our trust in going forward. I find that notion ludicrous.

I would hazard that the probability of a major news story breaking within the next year before the 2024 US Presidential election involving content produced with generative AI is greater than 50%, and that percentage will not drop at any point over the next year. We will have an event that causes a shift in public opinion, which will be wholly synthetic. The corporation that owns the model used to synthesize the hoax will disavow or deflect blame, but the damage will be done.

You don't need a literal Skynet to have an existential risk to life posed by AI. You don't need a bomb to go off when your target is standing in quicksand. All you need is hubris to believe that we, collectively, will always be one step ahead of the self-interested, well-armed, and ethically challenged few. The arms dealers, in this case, are profit-driven, unbeholden to the bulk of society, and similarly ethically challenged.

Back in August, a federal judge decided that works of generative AI aren't protected by copyright, and I agree. AI accelerationists (and AI companies) disagree. Discourse! By limiting the protections that AI-generated works are given, we bypass a lot of the economic concerns that people have about AI: We disincentivize models that scrape repositories of human culture without consent or compensation, and we remove profit from the equation of cheap, AI-generated content.

The philosophical answer as to why we can deny generative AI authorship is pretty simple, and I can make the case without stepping into the morass that the judge did in this case. His argument was there was no precedent to grant copyright to non-human entities. While this may be legally true, it sidesteps what makes 'authorship' different from 'content generation': intent and subtext. Generative AI has neither, at least to the degree that we can comprehend it.

Intent is judged by going back and asking the author what they meant by creating something - the why behind any creative work. When I write a poem, I (usually) have the subject firmly in mind, maybe some bits of inspiration I can point to. Sometimes, I can tell you exactly why I chose a specific word – whether it was the meter, sound, or meaning.

ChatGPT and Midjourney can do none of that. While it's possible to introspect the inner workings of LLMs, what you wind up with is a bunch of numbers that represent vectors. There's nothing to attach those vectors to, making the whole thing opaque. They don't have meaningful metacognition - and since that is true, I'd make the case that authorship is beyond them. Authors can explain their work to you - AIs can't.

If you don't have an author, you don't have an original work. You might have an incredibly convincing collage of parts that convincingly passes as original, but it is a simulacrum, a finely-grained average of smaller and smaller arbitrary vector choices. In some cases, this isn't even metaphorical - you can see the reality of generative AI if you get very, very close.

Midjourney recently released an upscale feature which is much more powerful than previous versions. The image size that you can produce with the 4x upscale is frankly breathtaking. I was working with some images, when I decided to put a few through Photoshop and inspect what to my eye in the originals was fine details - a filigree of lines. What I found surprised me.

Here we have the conceptual design of a board game, as imagined by the legendary artist, Möebius. It's pretty good as these things go - I've been using the new tuning feature to really hone the creation of these sorts of illustrations. You have a collection of 3D buildings on a board, at least one piece, and some spaces on the board that seem to have some commonality. Its at least a decent jumping-off point for me as a designer to think about how I'd implement something like this.

If we look very closely at individual elements, the style offers a peek at how the sausage is really made.

This is one of the buildings on the board. There's not a straight line to be had. Windows morph into corners, the base of the building is round at the base and square at the top, and the architecture falls apart into a sad birthday cake.

Here's another building with what, in the original image, appears to be a range of masts and towers for broadcasting or some other collection of technology. You can see, particularly at the top of the building where it overlaps the track behind it, where the AI stopped dealing with the building as part of the foreground and blended elements into the background, even molding the shapes to the blue track behind the object.

Famously, generative AI is bad at drawing hands. This makes more sense when you consider that there's no "planning" involved in AI art - there are simultaneous decisions that are made at increasingly smaller scales until you have hands like the above. Hands are things that physically get in each others' way - you must compose them, and there is no composer.

And, since there's no composer, the result can't be considered "a composition" - of word, image, or otherwise. Title 17, Section 102 of US Copyright Law, has this to say about works subject to copyright:

In no case does copyright protection for an original work of authorship extend to any idea, procedure, process, system, method of operation, concept, principle, or discovery, regardless of the form in which it is described, explained, illustrated, or embodied in such work.

I'd make the argument that generative AI is all process, procedure and system, and as such, the product of that process, even an intricate one, isn't protected by copyright. There may come a time when this isn't the case - but for the current deluge of generative AI - the porch light is on, but no one is home.

So maybe you've heard that AI is having a moment.

I want to be explicit about my own lens here: I am a developer, an artist, and a lover and avid consumer of art. I understand the impulse to make these generative models, the fears surrounding them, and even the lust to exploit them. The art for this site is all either AI-generated ex nihilo or generated using a photograph I've taken as the basis – a collaborative work if you will. We'll get into whether prompting AI is inherently collaborative in another article.

Having said all of that, in exploring these models you start to realize a few things that show that the majority of the power that people assign to AI art is actually human interpretation of that art, not some inherent quality of the art itself. It takes a fair amount of experimentation, but I will say with confidence that AI is not nearly as amazing at producing art as the human mind is at elevating it, often far beyond the status it deserves. We must realize this basic idea – the faster the better – so we don't despair over the new wave of technological disruption headed our way.

Less Is More

If you want to dig into the cracks in AI and see what it's made of, you give it very little to work with. What happens when you don't give AI a whole thesis on what sort of result you want, and just let it do its thing? Fucking weirdness. That's what happens.

(Note: All art in this article was generated by me using Midjourney v5.)

Always A Manic Pixie Girl

For instance, let's look at how Midjourney interprets a prompt that's just the names of the first four elements on the Periodic Table:

/imagine hydrogen
/imagine helium
... you get the idea

The first thing you realize about Midjourney AI is that it loves putting people as the center of focus of an image, often when it isn't called for. The next thing you notice is that those people tend to be women. Then you notice the age and relative attractiveness of those women. It is incredibly consistent:

The third element, lithium, gets a different palette of young women, one with Harley Quinn hair, but again – when in doubt, slap a feminine face in the middle and ship it out – probably it will still get a thumbs-up from the user. Of course, this feature of AI tells us more about the users of Midjourney than anything about the AI itself. Hell, I like Lithium #1 just on principle, and at least that has metal in it!

It's not all manic pixie girls, at least when there's an abundant source of association with the word. Helium gives us balloons, mostly:

Helium leads us to balloons, and balloons as a concept are related to kids as a concept, so the images as a whole tend towards some commonality with the word. Let's put a pin in that thought for now.

I'm not entirely sure what's going on with Helium #1 – it looks like Sugar Rush from Wreck-It Ralph. It's not until we get to beryllium, the fourth element on the table, that we have a different kind of face.

... and then we get acting legend and all-around awesome activist James Cromwell in an airplane hangar. It's not until boron that we get a batch of images that have nothing to do with the element or even the washing powder, but at least aren't four portraits of someone masquerading as a chemical.

This image does give us a glimpse of other gestalt nuggets rattling around in Midjourney: people gawking at something in the sky, random spheres, and not-quite-Harry-Potter. That archetype is another common figure that gets thrown into the mix.

By the way, each of the palettes that I've pasted above was the first result I got back from running the prompt the first time. The case doesn't get stronger for Midjourney if I make it think I didn't like what I got the first time around. After two more iterations of the helium prompt, I got:

Helium is a redhead. Who knew? And finally the sixth element, the noble gas argon:

We're not quite back where we started, but it rhymes.

Great Balls of Fire, In the Sky, With Diamonds

Glamour shots aren't the only thing that pops out of Midjourney when it doesn't know what else to do with a prompt. I've found that another go-to is the "tiny people in the foreground gazing in wonder at something in the sky", I've seen it several times before, but it struck me when I gave Midjourney the prompt /imagine archaea – archaea are incredibly ancient, incredibly tiny single-celled organisms that don't have nuclei – and I got this:

Or how about we just imagine a forest?

Or the word "dandriff", which is just an alternate spelling of "dandruff" – I'll admit that I've used some very strange words from this API in exploring AI art – gave me this piece, worthy of a dark-light poster in a weed dispensary:

To be sure, I have no idea where this motif comes from in Midjourney. I don't know enough about how image generation works to do more than speculate. I hypothesize that it has something to do with how the image is generated from noise – if you watch Midjourney working, images always start as a roundish blob of one color on a different background color. Maybe without any real guidance from the prompt, that blobbishness just gets refined and crystallized.

I Can't Even.

Then there are the images that come boiling out of AI like surprise eels from your toilet, and where wind up surprised or even disturbed by the alien-seeming logic of it all. For the sake of the general audience, I'll skip over some images that look right at home with H.R. Giger, or Eraserhead. These are more along the lines of "WTF?"

Midjourney's attempt at the visual representation of "Javascript":

Or the town of Muonionalusta, where a meteorite fell about a million years ago in a remote part of Sweden (my attempt here was to get a texture like meteoritic iron):

Or how about some nonsense words? That's pushing the envelope even further in terms of AI weirdness. For instance, the nonsense word "loab" grants us a giant orange owl cyclops over a floating island mansion.

Or the phrase "grink flurn", which gives us this palette, which includes at least one manic pixie girl:

Hey Midjourney, what do you think of technology? Oh ... ok.

And That's Where You Come In

The overwhelming question that comes to mind when I view most products of AI art is:

WHY?

What possible combination of factors, reinforcement by the user base, programming bias, or just plain stochastic chaos in the models creates these images that I find so perplexing? My mind immediately wants to know the reasoning behind the choices made by the AI, the image fragments it selected, and the endless loops of choices made.

It was while I was doing a slew of these sorts of prompts that the answer came to me: Asking the 'why' of AI is a mug's game. The AI is just doing what it does procedurally – looping through all of the combined neural pushes and pulls of the model until it hits the end of the run.

It's the person taking the image in that's struck by the combination of elements, and meaning is both extracted from it and imbued back into it. Granted, this is true for any artwork - I mean, at some level, that's what art is. Where things get tangled up is in our natural reflex to look at the output of AI and expect that it was created intentionally – that there's a message being conveyed with the images that go beyond "this shape and color from this image was tagged with 'dinosaur', so I'll mix it with this other image also tagged with the word 'dinosaur'". We give AI far not only more credit than it deserves for the composition of the images that we see, but we also give it agency. For lack of a better word, we give AI art soul.

That's the real crux of the issue: The only true intelligence in AI art is you.

That doesn't hold true when we look at human-generated art – the artist made any given piece for a reason, even if that reason is practical, obscene, or any other excuse under the sun. That can feel true for generative AI, but it isn't. There are influences, sure – there are plenty of documented cases of sampling bias in AI, for example, with professions. I feel that the majority of what AI art evokes in us comes from our own pattern-recognition processes fighting with what looks like a piece of human art – something made with purpose and invested with time, energy, and occasional talent. Instead, AI art is, at best, a lucky collage of visual noise. Our brains can't square that circle, the lack of soul, so we immediately start inventing the purpose and projecting that back onto the art.

It just isn't so, no matter how many people become convinced that current AI is sentient.

That may change in the near future, and believe me when I say that I feel that change coming straight at us. But right now, everyone needs to take a step back and realize that the power of AI art is in the human looking at it, not the generator. The powerful feelings we associate with generative AI are us, reflected back at ourselves. We're the only soul in the room.