It’s hard to escape the idea of AI these days, with models refining their approach and appearing more and more every day. ChatGPT, Stable Diffusion, MidJourney, and more – they all promise to be useful tools and they are the new tech grift buzzword that is hard to escape. Some bloggers I know have made content out of asking AI questions and repeating the answers with analysis. Some people I know use ChatGPT to write employee reviews, to synthesize songs in a specified style, and to create art they can use in some form later on as an asset.

There’s just one problem – most of it sucks, it exists in a legal grey area that is quickly being defined against AI, and even the use of the term AI is a fiction.

Generative AI is on the hotseat because of the potential it shows to executives looking to save a few bucks, looking to outsource writing from underpaid humans to a piece of software, to push concept art production from amazing artists to a tool built on interpreting those same artists’ work, and to create large pieces of a media project without needing more than a person to type a prompt and staff to clean up that prompt into something usable and interesting.

Admittedly, these tools get a lot of free and unquestioning press because they tap into that part of our brains that wonder about the limits of technology – fearfully or not. Can a machine think? Can we teach the concept of intelligence to a computer, to make it not just a logic machine that performs math until it escapes your eye but imbue it with the skills of reasoning and determination? Can we ask a computer questions of ethics? Can it be self-aware?

The answer to all of these, for now, is no, but generative AI coasts on the representation and idea of what the term “artificial intelligence” represents to people. It wants to be perceived as bigger than it is, this space of infinite possibility that represents a crowning achievement of humanity and the future of how we’ll all work.

So what is it?

They’re machine learning models. The initial seed is a massive input of data, connected sometimes with metadata like keywords and descriptors and sometimes, depending on the model, left for the machine-learning model to parse on its own. The ML model scrapes this data, makes observations about it (called inferences), and then stores those inferences as part of the training. When you input a text prompt into one of these programs, it calls those inferences and tries to assemble the trends it has learned about what you’ve just said into something. It doesn’t think about this, and there’s no logical process where the machine is capable of thinking about it in the way we would think of thinking. It just knows that its inference data points to certain texts, stylistic techniques, and so it grabs data and tries to jam it in to create a response.

This is cool on some level, in terms of achievement – an art AI might be able to give you a cool picture of a painted, brush-stroke heavy city skyline if you ask it to make a skyscraper painted like Van Gogh, but it can only approximate the style and the subject. It’s kind of neat, and it might genuinely look nice to a point to someone – but it’s also always missing something.

Worried about ChatGPT writing papers for students in college? Well, it can’t – I mean, on a technical level it can, but it doesn’t know how to properly cite, and even if told to cite, it will generate a garbage citation based on what it knows citations look like from the training data it was fed. The odds of that citation referencing an actual written work worth citing are 0%, but hey – cool trick. When people call these models AI, they want it to sound vague and fascinating, but also far more capable than these models actually are. ChatGPT can be bullied – if you tell it to tell you what 2+2 is, it will spit out 4, not because it is a logical machine but because its training data is loaded up with inferences of this fact. If you tell it, then, to tell you that 2+2=5, you can get it to do that – and you can bully it into a fake apology for getting it wrong and telling you 4 too!

All of this is crucial to understand because these models are not intelligent. They cannot think about the topic or reason on it – they can only take a text input and respond with a basic reply that fits with scraped inference data it has. It has been told how to do a thing and it does it a specific way based on that programming – it doesn’t deviate and it is incapable of thinking more about it. If you want a chatbot to give you a good text story, you have to feed it a prompt a mile long, and even then, that story isn’t quite going to be exceptional. It might be funny or charming in a way, but it is just the regurgitated product of millions of lines of text fed to and scraped by the ML model at the heart of the system. It is fictionally humanized by basic lines of response fed to it about how it should handle certain inquiries, so it can apologize, but it isn’t apologizing itself, it is just responding with the string of text it has been told to use there. Not to mention that the apology is itself hollow when you can put forward your inquiry in a role-play format and get it to answer anyways!

I recently spent eight weeks during a class researching generative AI as it currently stands and its use in video games, and that time has given me a lot to ponder. A lot of thoughts I had, arguments against it I couldn’t necessarily crystallize in an academic format, were given life by this post by game developer Doc Burford and it is worth a read. I want to start with the game industry specific stuff and then make a broader argument.

What Are Game Studios Doing With Generative AI?

So far, not a lot. But that is a very temporary answer, because it is coming. Blizzard (frequent topic of this blog that started and largely remains about World of Warcraft, imagine!) has made an internal tool called Blizzard Diffusion, designed to generate concept art for their games through an AI model. Ubisoft was using an AI tool to write scripts for them automatically. Coding addons for AI, probably a best-case scenario, are in use in Unity as an engine and are also becoming popular addons – GitHub offers their own version that works with a handful of integrated development environments.

Right now, these tools are promised not as ways to reduce headcount, but instead to increase output and quality. Blizzard Diffusion is supposed to be trained on Blizzard art and so it should, in theory, mimic the Blizzard house style pretty well, with the idea being that it should help the studio deliver more content on-time and keep their pipeline rolling.

In practice? Well, we’ll get to that.

What Are The Objective Pitfalls of Generative AI?

Right now the biggest issue is that generative AI falls into a dubious copyright grounds that no rightsholder ever wants to be standing on. Here in the US, the current ruling is that portions of a work generated with AI cannot be assigned to a rightsholder, which means that any character or world designs made with AI, any text or dialogue written with AI, cannot be held in copyright and thus, cannot be safeguarded legally. If you want to sell merch with your new AI character as the star, anyone else could sell merch with that same character. A big part of modern fan culture is that anything in the franchise should be monetizable – character likenesses, quotes, shots and art of the environment of your story – all of that is packaged and resold in derivative works in other media, consumer goods, collectibles, and the like. Blizzard here, to pick on them again, has made bank off of distinctive character designs sold in multiple forms. Diablo IV is a hit in part because of Lilith, and statues of her costing $600 are a thing that Blizzard is selling, and will likely sell out of.

If legal precedent continues to establish that copyright cannot protect generated art and text, then the likely solution for a lot of these companies is that they simply cannot use it because it poses a risk to their income streams for a non-copyright protected piece of their work to be out in the commons and with no legal recourse against it. I am not a lawyer, so there may very well be loopholes around that – but it seems like a pretty big issue!

The next is that AI models have a long way to go with regards to closing gaps with human artists. These models need a lot of training, and without the actual intelligence and reasoning capabilities to understand what it is seeing, AI models can only make vague inferences and attempt to construct new work based on what it has seen. It has a concept of the human hand, but it is trained on images where not both hands are visible, where hands are angled and partially visible or obstructed by other objects in the frame, paintings and models where the hand might be a suggestion of form rather than a detailed image, and of course, images of people with genuinely different hands – fingers removed for a medical issue, as an example. You can try to brute-force this by feeding the AI model images of hands for days, weeks, months, even years. You can have it parse through millions of hand images, but at the end of the day, hands all look a little bit different and because the machine cannot reason, it cannot fully understand why that is. So it will always output an approximation, one that can get better over time based on inputs, but it will always struggle at least a bit. The same is true for all sorts of body parts and things – feet, stuff like elbow and knee creases, and more. Without that crucial reasoning capability, these models will never quite get it 100% right, although they can be attuned to get closer and even, functionally, close enough.

Lastly, AI models rely on underpaid and overworked human labor and the stories that come from that are often quite bad. OpenAI, the makers of ChatGPT, used an outsource company in Kenya to help make ChatGPT less toxic. In a Time investigation published in January 2023, the outlet found that workers were paid between $1.32 and $2 per hour and were asked to review an insane amount of messages of disturbing content, with the deal collapsing after the outsource company was asked to source and review images of a disturbing nature, including illegal images of child sexual abuse, other forms of sexual violence, and disturbing graphical violence. The claims are disputed over the nature of the requests, but there is no dispute that around 1,400 such images were rounded up.

Without even getting into the more subjective discussion of artist’s rights and the ethics of AI on that front, there is little denying that a big part of the power that makes these tools work isn’t technology or magical models – it’s people, often underpaid, poor people in countries we scarcely think of, subjecting their eyes and minds to horrible content so that we don’t have to see it or think of it.

The Labor Case Against AI

AI is angling to take away jobs. Well, not the AI, but the people who would use it.

Most of the hype for AI is being built not by single evangelists, the companies responsible for making it, or some cases where it can be genuinely helpful, like solo game development or in-house tools trained to do one specific task to aid workers, like the machine-learning animation tool used on the most recent Spiderverse film – but by big Hollywood film studios and big video game studios.

Why? Well, simple – in the face of losing ground to exceptional labor movements, including the potential of an actor’s strike on top of the Writers’ Guild of America strike that is ongoing and the CWA allying with video game labor wanting to unionize, studios have AI as a back-pocket option they want to use to bring their workers to heel. You’ll work for what we want to pay you, or else be replaced.

Obviously, this case is fraught, because the copyright situation with AI is genuinely something that could torpedo this strong-arming, but it is also early days.

For the major players in entertainment, the idea of AI is intoxicating precisely because it enables a dystopian automation future – only the thinnest of cleanup crews to fix up AI work before spitting it out onto the screen/hard drive. The bet is that AI-generated slop entertainment could still bring in money, and as the technology improves, the quality of the resultant work hopefully goes up too. The sad thing is that to a point, I think this would be successful – given enough tweaking by a human staff, AI-generated work could maybe trick some percentage of the population. There is a question of where the line is – a video clip I have saved for later in this post will reveal that AI-generated video can be pretty bad in terms of quality, and there are layers to the problem like how much you involve the AI (what combination of writing/voice/video), but for some people who just want to be entertained, it could work. That is, perhaps, commentary on society, but that’s the bet, and I don’t think it is altogether impossible.

The Subjective Case Against AI

I think AI work is shoddy, gaudy, and often lacks any form of interesting cohesion.

My guild has made a small tradition of celebrating milestones in Heroic raiding with stories about raid NPCs and our raiders written via ChatGPT. It’s funny and quirky, but the stories aren’t deep and they often require our people feeding the prompt to lead it on a journey to a destination, instead of just letting it roll. The stories aren’t important, but they also aren’t good – they elicit a chuckle, we all laugh, and it is largely forgotten except for weird turns of phrase or the other quirks of AI.

Visual art in AI has this overprocessed look about it. They look like Lisa Frank art devoid of the themes and central cohesion that those works had, and often the people making them paint over them in Photoshop or other image editors later to try and paper over the deficiencies on display. They can look cool in a way, but it is a surface-level cool – there’s no larger thought or merit underpinning it, just the ideas fed in via the prompt. The infamous “Hollywood is done for” tweet showcases one of these stunningly bad works with a person who is, believe it or not, proud of this thing!

Hollywood is done for – you might not believe me, but this is AI generated. pic.twitter.com/2cV7BWeVdc

— Chase Curtis 🍭 (@realchasecurtis) April 29, 2023

AI art struggles for me because it is just-so – it follows the prompt like a machine would, based on a well-laid set of rules and constraints, and so it lacks something…

The Use of the Word “Art”

Here’s where I get a little saccharine, so excuse me for a moment, but I think this is a crucial point to make.

Art is human endeavor encapsulated in a given medium. We take our experiences, the little pieces of our lives that make us unique and give us meaning, and we imbue them upon a song, a writing, a painting, model, game, movie, play, whatever the case may be. When we as humans create art, we do so with intent – the intent to say something, represent our lived experiences, make people laugh, cry, or just generally feel. Art is such a visceral and real part of our human lives that touches us precisely because it holds a mirror up to us and asks us what we think, how we would handle things in a situation, what we would do. It often asks us hard questions we hope to never have to answer in real life – how to handle a shootout, being the last human alive, the last bastion of hope against a demonic invasion – and it can put us into situations where we do have lived experience and show us different ways – romance stories, adventure and travel, even mundane stuff like cooking TV.

Art gets to be “art” because it has intent and soul behind it – even the worst, corporate assembly line art has a message and is imprinted with the ideas and experiences of the people behind its creation. AI art lacks intent – it fulfills a basic prompt and nothing more. It has a void at the heart of it, an empty maw gasping for some deeper meaning and finding nothing. You cannot write a prompt to communicate your feelings by having an AI generate a thing – it will only ever be a clinical interpretation of what the thing you asked for is.

This might sound snobby or even weird, but I will always prefer a basic, technically-bad piece of art from someone who tried to make it themselves over a pristine, technically-proficient AI-generated thing. The AI-generated content might look more detailed, more full, but the basic, bad art from a person not good at the process will at least carry feeling, meaning, and depth to it, and bear the imprint of the person who created it. I’ll take the no-training music of Nobuo Uematsu over someone telling a music AI to generate something that sounds like Final Fantasy VII music anyday, because the AI has no intent behind creation, feels nothing about it, and cannot have pride in the work or any greater thing to say – it simply tears through its inference library for Final Fantasy VII data and slops it into an MP3 file.

Some people I see arguing in favor of AI point at stuff like music sampling as a way in which humans often also rip other works or how artists mimic the style of their favorite artists and acquire learned traits they like to reuse, I still think that humans do it better, because even the artist who really likes Van Gogh might use his style in a new way and they will imbue their work with their own identity, and musicians sampling are using those samples like a plate to build a new meal upon – sure, some of it is familiar, but there is a new piece of art and a new message built upon it.

AI work is not art to me because it lacks that human essence that makes art feel transformative and interesting. It can be technically proficient and appealing on some level, but it’s not art (to me).

I think a lot about how AI writes because of those guild ChatGPT stories and the thing that always stands out to me as someone who writes (and inexplicably has readers, imagine), is that AI text has no voice, no unique viewpoint. As I’ve developed my craft over the last 7 years here (and longer in other forms outside of this blog), I’ve always centered the idea of having a voice as a writer. For me that means writing conversationally, it means breaking grammar rules when I think it makes a sentence more interesting or engaging, it means abuse of em-dashes, and it sometimes means particularly bad writing habits for the sake of style, like a massive paragraph-long run-on sentence. ChatGPT can’t really do any of that, not with meaning – it might parse some text and rip that style off, but it doesn’t know why it did it and it doesn’t do it with any intention or purpose – it’s just a thing it saw in some scraped data.

And now we should talk about AI bros.

If AI Is Your Solution To Most Problems, You’re Hopeless

AI bros love talking up AI as the future. It follows, to mirror Doc Burford’s observations from the post I linked early, the NFT and Web3 movements. You start by promising this big new thing that is revolutionary, and then implement it. People start to realize it actually kind of sucks and doesn’t do the thing you claim it should, so you get sour and start promising it is the future, strong-arming people and being belligerent as the obvious deficiencies of your future become apparent.

Calling these models AI is a misnomer because it implies a far greater degree of thought and process than most of these models can do, and it sucks in a certain kind of rube that doesn’t critically think about why it works or how – just that it spits something out and that something is good enough. If you need ChatGPT to write, you’re not a writer and if you lean on it every time you need written content, you’ll never be a writer. If you need programs to spit out art, music, voiceover, video, any form of content – you’re not a creative and you damn sure aren’t an artist or experienced in that discipline, you’re just a parasite leeching off the work of others fed into the model. I’ll say that and stand by it.

The biggest problem I see is that a lot of AI bros, just like NFT bros and Web3/crypto bros, buy into the hype in a way that makes them blind to the issues. NFT bros genuinely believed that a blockchain mortgage was going to be a good thing because it would be forever on the chain, even though that doesn’t solve any existing problems and actually creates more issues because the legal status of a mortgage is recognized based on existing infrastructure and designs and would require layers of legal process that, frankly, most Web3 assholes don’t want (if you have to report a home deed or mortgage on the blockchain to the government, that opens up a lot of vectors for crypto to be regulated, visible, linked to an identity, and taxed!).

I know an AI bro, one, and his solution to everything is to plug it into some stupid AI model, to a point that he doesn’t even think criticically for himself about the things he asks ChatGPT. He assumes that ChatGPT can do reasoning and actually math things out, as opposed to being a chatbot that can spit out a text response and only responds based on text inputs it has stored from inferences. Asking ChatGPT about the drop rate of the Evoker legendary is really dumb because the model isn’t built with that data (the question is based on info that is too new!), it cannot analyze it with any real accuracy (because it isn’t built to do that!), and it tells you what anyone with basic reading comprehension could have even if it does get it right (since you get one “roll” per difficulty, with a first kill on higher difficulty giving rolled-up chances for the lower-difficulty raids as well, then the efficient way to farm it is to kill the highest-possible difficulty, yeah). So basically, you wasted time and a prompt on a really basic, easily-solved logic question that it can’t even actually do since it doesn’t do reasoning or math and just spits out text from the banks it has scraped. Cool!

At a certain point, taking pride in the stuff AI generates for you is almost insulting, because you didn’t do shit. Don’t show me your ripoff FFVII MP3 music track you had an AI make – I am not listening to that because there is 0 chance it is anything worth my time or anywhere close to the artistic value I got from the tracks made by Uematsu. I have 0 interest in reading your AI-story of more than 3 paragraphs, or looking at your AI-generated 3D model. I certainly don’t want to play a game made entirely of AI-generated assets and content.

And I think that touches on the last big thing I find irksome with AI and modern culture in general. (epic boomer moment)

Art and experiences we have with it are not “content.” I’m guilty of this in the context of modern games too, but “content” is such a distant and awful way to look at art. I think the Marvel movies have become popcorn trash, but they are still art (I do agree with Scorscese here a little bit, though, don’t @ me). I think modern WoW is gameplay-first in a way that discounts the artistic value of it a little bit, but it is still art. I love pro wrestling, and it is often trashy and weird, but sorry – still art.

Where I think it gets bad is when people creating the art refer to it as content, and I think that mixup is where AI bros so often thrive. They think they can make content (they can’t) and could fill an entire media project with AI-generated garbage and have it result in something worth experiencing, and they are wrong. Yes, you can make a heap of assets and content with AI, and you can assemble it loosely into a thing that resembles art, but it isn’t and will never be art. The more you lean on tools to do the work for you, the less chance you have of developing those skills or honing your craft – you’re just going to be stuck relying on AI forever, and that isn’t a good thing to do or good position to be in. It reflects a lack of desire to learn or do better to lean on AI to make stuff for you over learning the skill yourself or finding someone with that skillset to help you out, and this lack of care is reflected in the end work. At best, your AI-generated thing will be a cheap facsimile of something better that came before, that was made with intent by a human creator channeling their lived experiences into art.

I’m sympathetic to people like, say, solo game developers without a budget to hire from, but I’d almost rather an asset flip than an AI-asset game, because at least assets you can buy or get for free in slow-motion via the Epic Store are made by people with some measure of artistic intent, and if you slap well-matched assets together with a post-processing shader or unifying visual effect, it can look good and play well! When people make things as a proof of concept for what AI can do, I always find it lacking in some way and I think that we can and should expect more of our media than what AI bros want it to be.

In short, my position is this – generative AI isn’t even AI, what it can do is limited, its use in media is a mess of legal and ethical questions for objectively lower-quality work, and only those who are fundamentally incurious simpletons will express amazement at and reliance in AI, while the rest of us mostly just sigh. It has genuinely bad possible implications for creative work, and the future most AI bros profess to want kinda sucks. It’s not the future and relying it today doesn’t make you forward-thinking or ahead of the curve.