Words: 10549 Approximate Reading Time: 1-1.5 hrs.
The rise of AI (or what we call “AI”) over the past few years has brought with it a host of debates concerning its functionality, its usefulness, and its overall value. It’s not hard to find people sharing stories about how they created something – a video, a “drawing,” a book, a program – mostly or entirely with the use of AI. It’s equally easy to find people sharing stories about asking these an AI chatbot a simple question that is answered with complete nonsense.
The companies behind these tools have a need to make money. Ungodly amounts of money, if possible. And that need compels them to push these tools into every facet of human existence. Similarly, businesses and managers see the promise of AI as a tool which can radically increase the efficiency of workers, meaning that they feel compelled to adopt AI. As such we find AI continually seeping into our lives, regardless of whether we want to use it or not.
Steam’s most recent Next Fest – an opportunity to highlight demos for new and upcoming games – raised a lot of eyebrows because it featured somewhere between 7000 and 8000 demos, nearly double the most recent Next Fest. And much of that was due to the churning out of games made primarily or entirely with AI – very often low-effort clones meant to check certain popularity boxes to get people to pay attention. Epic Games recently released a new version of its online store/launcher, entirely created with AI coding, and is pushing its developers to use these tools exclusively.
There are too many angles of discussion about this topic to handle it holistically within this space. Instead I am more concerned with the question that I think all of these discussions fall back to: should we be using AI or not?
The core of the question is sincere. For some, AI is simply a tool. It accomplishes a task, and it might not be perfect, but it might improve over time. That tool allows people to work more efficiently than without it, so there’s not much of a reason to avoid it.
The argument against tends to go all over the place, in part because it is composed of a lot of people from different backgrounds with different concerns. There are environmental concerns, artistic concerns, economic concerns, ethical concerns, and political concerns. And because of that jumbling, it can feel like one concern gets picked and held up as the Ur-Problem.
The scope of this all being too much, I wish to focus on a few key points. But I want to make a case for the destructiveness of AI. Not that AI is something to avoid because it is an inherently evil tool, but that it is a tool wrapped in layers of political and economic baggage which makes its use unethical. Perhaps there can be a time when its use is innocuous, but that time is not now.
Defining Our Terms – AI and Anthropmorphization
One of the central problems in discussing AI is the term itself.
AI is Artificial Intelligence, and is a common element of science fiction stories and pontificating about the future. The concept revolves around a machine that can think in the same way that a human thinks, though often on a more powerful level. Where machines merely spit out results based on inputs, an AI should be able to understand those inputs. You cannot ask a calculator a question, but you could ask an AI.
The problem we have is that the thing we are currently dealing with isn’t AI by that definition.
Current AI models are built off of machine learning algorithms, primarily Large Language Models. These algorithms are designed to pull together massive amounts of training data to establish patterns, and from there use those patterns to recognize inputs and create outputs.
The common analogy, and one that is still apt, is the word suggestion on your phone. As you write out a sentence, your phone suggests a set of words to select for what might come next. Sometimes that word is what you’re looking for, sometimes it isn’t. The words are selected based off of a probability model: based on the algorithm’s data, it has found that X word is most commonly followed by these three words. It might even adapt to your own style as you type, recognizing that you tend to follow X word by something else and try to suggest that more frequently. Large Language Models are much more powerful and sophisticated, but that probability model is still important.
Because what’s key is that there is not an actual understanding going on underneath all of this. Your phone doesn’t know what you’re trying to write. It’s not trying to be helpful. It has been created for a specific purpose, and it is simply producing an output based on its design.
Hence one of the major pitfalls when discussing AI is attributing thought or intent to it. Which is not aided by its own output. Interaction with an AI often produces what feels like a conversation, and the responses we get feel like there is a mind behind the screen, even if we can’t see it. Which makes it harder for us to not see the AI we use as a true intelligence, almost as a person.
But there’s nothing there. We often see stories of people claiming that an AI model has achieved sentience, but the basis for this tends to be the person simply taking the outward appearance of things for the truth. They see a machine that has been programmed to talk like a person, and thus presume it is a person. No less than mistaking an image of a pipe for the real thing.
The use of the term “AI” is by no means accidental. As I already mentioned, AI as a concept has already existed within the public consciousness. And as such we have a general set of feelings about it and what it can do. A true artificial intelligence would be extraordinary – and the companies developing these models are aware of that history and the value such a true AI could bring. In short, “AI” is itself a marketing term, a way to siphon money from investors and lure in customers to use it. There is far less profit in marketing these as mere chatbots and algorithms.
The prevalence of the term sort of forces our hands. Even though the thing we’re talking about isn’t really AI, it also has cemented itself as “AI.” It’s hard to not use that term now. Often people will use “Generative AI” or “genAI” to try and specify the particular models that exist to generate code or writing or video or images through this probabilistic method. Insofar as “genAI” is largely the subject of these debates, though, the term doesn’t seem to serve any useful purpose except perhaps to distinguish it from the idealized AI that has existed in the public consciousness.
It also creates other problems when discussing existing mechanisms. For example, video games have for decades had “AI.” This was just a term to refer to a set of commands for entities within the game: a complicated set of inputs and outputs. And yet, it is now hard to separate that term from this new breed of models, even when they have nothing to do with each other.
Nor is this helped by the term being associated with just about any kind of machine learning or algorithm. If your image editing software or word processor has a function that can detect issues, fix problems, soften edge, create sentences…at what point are those things “AI” in the same way that a model like ChatGPT or Claude is? Without knowing the way these programs operate, it isn’t possible for us to know which is which. And yet, the universalization serves the interest of those wishing to push AI – if you perceive everything to be AI, then it’s inescapable. You might as well just use it all.
The thing that generally has people up in arms, the actual focus of this all, are these specific models and the companies surrounding them. The confusion about AI allows these companies to hide within that gray area.
So I will continue to use this term just for the sake of simplicity. And yet, when we see the term “AI” we should be skeptical. We should be holding in our minds the fact that there are people who want us to make this association, and also that the association is incorrect.
AI Functionality and Error
The first question that should be asked is whether AI works. Which requires asking what it’s supposed to do.
Arguably, AI has two primary functions. The first is what it literally does: it creates something. You could ask it to write a novel, or make a cool image for your desktop, or make a video to share on the internet, or even program a video game. The second is its purpose: efficiency. By creating these things AI should allow its user to accomplish tasks more quickly, thus freeing them to complete more tasks or do something else.
On the first function, does it work when it’s trying to make things? This is a bit complicated because there are still distinctions in “making something.” Because something can be made well or poorly. If AI only produced bad books and images and such, would we say it works? Likely not. We can only say it works insofar as it’s able to make things that are good – or at least appear good. On a purely surface level, AI works. You can ask it to create a piece of writing or art or video or whatever, and it will do so, and that product will usually look right. Maybe there are issues in the details, but if you can spot those issues you might be able to request the AI spit out a product that is similar with those fixes.
As an example, I could have prompted an AI to write this whole essay for me. It could have produced a fairly long article, and there would have been some problems. The language likely would have been extremely empty, since it would simply be copying text as written by millions of others that it has scraped from existing books and essays. The arguments would likely have been weak, because making a good argument requires an understanding of structure and having an intentionality that the AI can’t have. I could have then made a request for the AI to touch up parts here and there. The end product probably would seem coherent, but would be lacking in value – it’s just a bunch of machine-generated text based on perhaps a vague prompt. It could have produced an essay, but not a good essay.
Now this failure tends to apply to things which are fundamentally “artistic.” You can ask an AI to make an image or movie, and it ends up with similar problems. AIs don’t understand the composition of scenes, so the videos generated by them often have strange and sudden cuts or contain dialogue that doesn’t make sense. AI-generated images might feel strangely flat, or fall apart as you start to investigate the background. As long as they don’t look too closely it can fool a viewer, but under any scrutiny the product falls apart.
However, more technical tasks which have outputs that can be verified are…well, the AI doesn’t always work, but it can produce results that work on a more reliable basis, at least according to many users. So for example, coding. Insofar as you need some lines of code that perform a specific task, an AI can usually generate code that fulfills that request. And as long as the user is someone who is familiar with the code in question, they can review it to make sure it looks correct. Likewise with anything else where a user-as-expert can produce something that can be clearly validated. Because these products tend to have more clearly defined “right” and “wrong” results, the results are more likely to be “correct.” These kinds of tasks are more narrowly defined, but for that subset of tasks we can arguably say that AI works in this sense of making something and making it “well.”
This artistic/technical distinction in what these AI products can and can’t do well is not a strict dichotomy. Some tasks are technical and yet require a degree of intuition or understanding that an AI might not be able to replicate. Or a task such as producing code might have a variety of “correct” answers (i.e. a variety of ways to write code to accomplish a specific task) which might be better or worse, and which an AI may not be able to distinguish. As an example of the latter, an AI might not be able to know if a more parsimonious solution exists, or might create variables that are difficult to track and understand for someone new on a team to grasp. So even if these technical products work, there may still be issues that are not immediately apparent. Such issues can be dealt with, but again they require the user to have an understanding of what they need so that they can properly review the AI’s output.
The second function of AI is to allow us to complete tasks more efficiently. Let’s use a technical task for this function. You want to write a series of programs: some of them are going to be fairly simple but tedious, and some will be more complex and may require some research and critical thinking to put together. Moreover, assume that the simple components are ones you have put together multiple times before, so you are already familiar with what is needed. In a scenario like this, you might use an AI to create the simple components for you, freeing up time to let you tackle the more complicated elements.
That’s the basic theory. The question is, does it hold water? The issue here is that there’s not a clear answer. There’s plenty of anecdotal evidence from various people claiming that using an AI saves time by completing tasks for them. And there’s likewise anecdotal evidence of users claiming that using an AI increased time because they had to spend more time checking the AI’s product and trying to get it to fix issues than they would have simply completing the tasks themselves. Broader studies have turned up similar problems – some studies show that productivity improves with the use of AI, while others have shown a decrease in productivity or no benefit.
I want to get into this topic more later, but the underlying problem of this efficiency function is that there is a distinct possibility it can only be fulfilled under specific circumstances. That you need to have a technical task, it needs to be a task that is relatively simple, that it needs to be something that you as an expert can validate very quickly, where errors are easy to spot, etc. Yet this is not the promise being sold for what AI can accomplish. These AI models are being presented as tools for every problem that can be used by anyone. Everything that you could possibly want to do can be done faster by using AI, no matter what the task, no matter who you are, no matter what your skills are. Does the product need to be carefully checked? That’s not part of the pitch. Will the product need to be revised, either by hand or by making another request of the AI? That’s not part of the pitch.
To talk about the function of efficiency we have to discuss it in terms of both what it does and what is claimed about it. In a strict sense it might be able to help certain users work more efficiently. But if for many others it leads to sloppy work or additional time use – regardless of whether that is a problem of the AI itself or “poor use” – then it’s not really fulfilling that function.
And on that note, let’s talk about hallucinations.
“Hallucination” is the term that has been applied to when an AI just makes up something out of whole cloth. If you ask it to write an essay for you, it might just provide some claims that are completely untrue based off of a source that doesn’t exist. There are plenty of examples of this happening in low-stakes environments, like people posing random questions to Google and trusting the AI response. But more serious cases have occurred, such as researchers citing fake studies listed by an AI, or the far too many cases of lawyers getting in trouble for filing legal briefs that contain completely fabricated court cases.
The threat of hallucination exists because of that probabilistic construction of responses. The AI doesn’t “know” that a given piece of information is real or fake, and it doesn’t even “know” if a given sequence of words is being pulled from something that exists or is simply being generated. So to take the legal examples, if you ask an AI to write a legal brief for you, it’s going to look at its training data to find legal briefs. And the data will have a bunch of claims followed by citations that follow a particular format. And then the AI puts together a response by generating a claim and then adding a citation to it. Namely, it will provide something that looks like a citation, but whether it’s pointing to a real case or just putting together words is going to depend on what the model’s coding says is the most appropriate word or phrase or term to follow up with.
It’s easy enough to chalk up these errors as something that needs to be checked by a human. That you shouldn’t be trusting the output directly. And indeed, in all of these cases you shouldn’t. Yet if we look to the way AI is talked about, particularly by the people selling these models, there’s very little discussion about these underlying problems. They aren’t ignored, but the magical nature of AI as something that can produce whatever you need is always the focus, and the need to double-check outputs is effectively an afterthought. People shout “AI can do all this cool stuff,” and then whisper “but be careful because it might be incorrect.” To write these cases off as nothing more than user error turns AI hallucinations into a kind of “No True Scotsman” problem – no true AI user would make this mistake, and so any problems that arise from hallucinations can be safely ignored.
And so we need to fold this all back into the discussions about these functions. Because we should be keeping in mind an AI’s propensity to produce these hallucinations – which are part of how the model works and thus not something that can simply be programmed out – and how those hallucinations impact both the product and the ability for people to work around them. In terms of the product itself, hallucinations obviously cannot be said to be “good.” If I have to engage in this thorough checking to make sure that the AI has generated something accurate, then it is only making something in the strictest sense that I now have a series of words or lines of code. We are back to the idea that the AI is producing something, but not something good. In terms of efficiency, perhaps I can at least take that product and fix it in one way or another, but the more effort it takes to check and fix the product, the less efficient the process will be. Which doesn’t mean that AI can’t possibly allow a user to be more efficient. Rather, there now needs to be a careful analysis of the tradeoffs, and it may well be the case that this analysis pushes us closer to the conclusion that these tools are only useful in a very narrow set of circumstances.
There are other angles from which to analyze these questions. But I do intend to end here with a claim that AI is being sold to us as more than it truly is. That insofar as we can identify these two primary functions, the AI doesn’t really fulfill those functions – at the very least, it can only do so under specific circumstances. That discrepancy between what it can do and what is promised is something I intend to look at later, but for right now I want to leave us with that conclusion.
The Economic Turmoil of AI
Whether it works or not, AI is one of those things that gets pushed on people aggressively. Companies have devoted an ungodly amount of resources to gathering data for their models, to construct data centers to house the processors to operate the models, to buy up processing power, and to keep those processors cool. And all parts of that process have an impact on the wider public. With that much money going in, those companies often need people and other companies to be using their products – in many cases operating at a loss with the intention that eventually they reach a tipping point in the market where they can become profitable.
And of course since this is the wave of the future, many want to ride that wave in one way or another. Some people try these tools out, find them useful, and then keep using them. Some of those people decide to become evangelists for this new technology. Sometimes because they truly believe it is revolutionary, but sometimes just because they can parley that evangelism into its own money-making opportunity. Some people fear that they could get left behind if they don’t adopt this technology, and so insist that everyone needs to jump on board. An untold number of stories can be found about some office employee being told that they have to use AI as much as possible in their workflow, regardless of whether that use is actually helpful, or even with the consideration that it could come with problems. See again Epic Games directing its coders to use only AI for coding, as well as other stories about companies pushing AI usage.
That push, however, does not simply end with some office employee now asking ChatGPT to draft emails or a software developer having Claude write their code.
Efficiency – one of those primary functions of AI – is meant to reduce the cost of creating new products. And an increase of efficiency in this respect means a decrease in time needed. So what happens with that extra time? Say, for example, that before you completed your work within a standard 8-hour workday. Now with AI you can complete all your tasks (and we’ll assume that they are of an equal quality) within 5 hours. And the same is true of your coworkers, as well. What is the response?
The ideal would be for that time to be returned to you. Now instead of working 40-hour weeks, you and your team members work 25-hour weeks. You all get more time to yourselves. Indeed, this is the true promise of artificial intelligence and automation from the communist theory of Karl Marx. Where the core work that is needed for us to survive – make food, build shelter, etc. – can all be automated with little input from all members of the community, allowing everyone to devote their time to tasks that they find personally fulfilling: art, crafts, writing, and so on.
Probably less ideal but still good would be for your employer to ask you to perform more tasks. After all, you now have more time. But in return your wages increase a commensurate amount. If theoretically you’re now able to do 40% more work than before, then you also make 40% more. There are economic reasons why this wouldn’t be good, but from a simple standpoint it at least would feel good to get a significant increase to the amount of money you’re making.
These are potential positive outcomes from the adoption of AI…but we don’t see any of these things happening. Indeed, we don’t really see any significant discussion of these outcomes. At best you might see these ideas shared in various corners of the internet, or that at some nebulous point in the future if AI becomes powerful enough we may not need to work at all. But the people pushing AI and the companies adopting it are not selling this all as a way to actually decrease the amount of labor needed. Employers are not suggesting that AI will be valuable for employees by letting them work less or earn more. The promise is always about “productivity.”
Instead, the outcome that we are generally seeing is that companies trying to adopt AI are hoping to reduce their costs by reducing the number of employees. If two employees with the assistance of AI can now perform the job that would take three people…why not just remove that third person? Insofar as there are or could be any true value from the usage of AI, the value does not get shared with the everyday person. Those people instead get to live in a state of uncertainty about whether they are replaceable by an AI. And those who find themselves replaced get to compete with all of the other people similarly caught in this same downsizing.
There is also a potential long-term problem lurking behind any downsizing efforts. As more and more effort is offloaded to AI models and the work of many employees becomes checking and correcting AI output, the actual work has to be done by senior employees with significant experience – the kind of experts who can reliably spot problems. And yet, a system built upon the backs of experts invites the question of where those experts come from. If, for example, a business employs only expert programmers who review code produced by an AI, it can subsist for a while. But as those employees retire or move to other jobs, the business needs to pull in new employees with similar skill sets. Yet those skill sets are built in part by the actual underlying practice. Coding, writing, analyzing, and so on all require a person to exercise the skill in order to develop it, and not simply spend time looking at outputs. So without copious opportunities for people to cut their teeth, there are fewer and fewer opportunities for people to gain the skills to fulfill those senior roles. Couple this with the relatively recent trend for many businesses to post hiring ads for “entry level” jobs that require prior experience – sometimes several years – and you find a system that is blocking basic skill development, allowing short-term gains (keeping and hiring experienced people at cheaper prices) to lead to long-term detriments (a potential absence of experienced workers after a decade or two).
The purpose of this section is not to say that the entire economy will collapse in 15-20 years. If this dearth of experienced employees does occur, there will likely be various methods of adjustment, and likely some pains that come with that adjustment. Rather, it illustrates the fundamental problem that lies with the push to adopt AI. That the adoption and all of its promises do not exist with the purpose of actually helping people. Instead it all exists to extract as much value from people as possible. You as an individual are not a person who uses AI. You are the tool that makes the AI work. If the AI could work without you, you can be removed from the process.
The Politics of AI
While we often talk about AI as just a tool, it’s important to remember that everything exists within a context.
Some aspects of this contextualization I’ve already discussed above. The way in which we talk about whether AI “works” has to take into account what has to be done to get the output that you want. And even then, merely getting an output that looks right shouldn’t be sufficient. The promise of what AI can do for us as workers has to be understood in light of what is actually being done by implementing this AI. It’s not being used to help the everyday person earn more or get more time, but to help those who already hold money and/or power get more money or power.
Political implications don’t need to be intentional. Say for the sake of argument that the outcome of widespread adoption was to weaken the ability for everyday people to hold jobs and push for better wages and working conditions. That AI makes so many people replaceable that people have to fight tooth and nail to compete over a dwindling number of jobs. A system that continually pushes more and more people into poverty. That outcome doesn’t have to be intended by any given person to be bad. A series of decisions by various people that seems rational on its own can accumulate to unethical outcomes. So when I discuss political implications of the widespread adoption of this current form of AI, it does not help us to get caught up on whether people specifically are trying to make any negative outcome happen. Where there is evidence that these outcomes are intentional, it is bad. But the even if there is no evidence (or evidence that such people have good intentions), it’s still bad because it hurts people.
These AI models do not pop into existence fully formed from a collective desire for neat stuff. These models are crafted by human beings, and anything crafted by human beings carries the biases from those human beings.
There’s been a significant amount of discussion from various corners about the problems of algorithms in various aspects of the modern internet. Social media – to take probably the most prominent example – ends up being designed in ways to promote continued usage. TikTok or Instagram or Bluesky want you to stay engaged on those apps/sites. And as that continued engagement turns into addiction it can have deleterious effects. An algorithm designed to feed you content aligned with what you’re believed to want can cause you to spiral into conspiracy theories. An algorithm designed to get people to watch and comment can push negative content into peoples’ feeds because that tends to get more engagement, fostering negative sentiments about all sorts of views (and effectively suppressing more positive or informationally valuable views and opinions).
And let’s not forget the feedback loop in play. As people start to understand how an algorithm works, they can also start to create content that feeds into that algorithm – to make videos or posts that are more likely to get engagement and be shared by the algorithm.
AI inherits all of these same problems, because they also work on many of the same principles. These AIs are not thinking entities (as I’ve talked about above). They’re models that take massive amounts of training data and use that data to figure out what output to create based on a given input. And models exist in part by having “weights”: variables that tell the model to prioritize or deprioritize certain inputs and outputs. Weights aren’t bad – they exist for all sorts of purposes to help make a model work. But those weights are also a way for humans to put their thumb on the scale.
So for example, an AI model can be programmed to spit out specific outputs. As a hypothetical, let’s say that you wanted to draft an essay about the problems of AI – all the different ways in which it could be harmful and why we should avoid it. Now maybe the companies selling the various AI models would prefer that bad information about AI doesn’t get spread. At the very least, they want to make it harder to find. So the companies program the AI so that whenever it talks about AI, it always emphasizes positive things. So when you ask your AI model to help you draft that essay, it keeps spitting out all sorts of good things that AI does or can do. If you’re someone who is already skeptical about AI (which you probably would be in this hypothetical, since you want to write an essay about how AI is bad), then maybe you’re already aware of this potential problem, and know what to make of the result. But if you’re not as familiar with how AI works, this result would likely be perceived as a kind of “truth.” If you’re now just a random person who wants to know whether AI is good or bad because you’re not sure, you might turn to an AI to provide various pros and cons. You would do so under the assumption that AI is fundamentally neutral. It’s an artificial intelligence, after all, it’s supposed to be neutral. And yet you would end up with all of this information and argumentation extolling the value and virtue of AI.
So a fundamental problem with AI is that as it becomes more and more widely used, it becomes a central source of information – a central source that is controlled by a small number of companies. Companies that are not themselves neutral. And by that I do not simply mean that they are likely to want to weigh in on debates about whether AI is good or bad. The people in charge of these companies have incentives to engage in politics: to favor particular parties or candidates all over the world. Sometimes that incentive may be personal preference, sometimes it might be an economic interest, and sometimes it might be ideological.
While an outlier in this regard, I think Grok – the AI developed by xAI (owned by Elon Musk) – serves as a useful example of all of these problems. Being controlled by a man who has a lot of very strong and very bad opinions and ideas, Elon Musk has a tendency to use his various businesses to impose his views upon others. Purchasing Twitter and connecting Grok to it brought a lot of opportunities for Musk to put his thumb on the scale of conversations. Perhaps the most prominent was the time when Grok began spouting off talking points regarding “white genocide” in South Africa – a pet theory of various white nationalists and racists. Grok wouldn’t just provide these claims when directly asked about South Africa, it would connect this topic to many other prompts provided to it. And this kind of output on a consistent basis cannot be a mere accident. It would require someone to make a conscious decision to program the model in a way to make these claims. Or we can point to directives for Grok to downplay ideas sourced from media institutions and effectively bolster ideas found from “controversial” sources (read: conservative, fascist, and/or racist publications). More simply – and maybe more amusing – are instances of Grok producing effusive praise of Musk, or directives for the AI to check Musk’s opinions on certain topics when producing its outputs.
Again, it’s easy to look at these examples as exceptional – Musk as a particular owner of a single AI model is abusing his power in this regard. But we should be taking two conclusions from these examples. Firstly, that there are still a lot of people who use Twitter both directly and indirectly as a source of information. Not only does it still have a large number of users, but it’s common for tweets to be shared in news articles, on Reddit, on other social media sites, and so on. Even if you don’t personally use it, the information that originates there can still cross your eyes. And many people will trust that information. Which in turn confers a lot of power to its AI. Many Twitter users turn to Grok to answer questions directly, once again under the assumption that since it’s an “artificial intelligence” it must be a neutral entity trying to provide a true account of everything. And so any manipulation of Grok is a manipulation of the information on and coming out of the site. Secondly, and more importantly, is that this is a threat that is true of any AI model. ChatGPT and Claude and every other AI model can be manipulated in this same way to direct its outputs. And those manipulations do not need to be as overt as described above. Even tiny nudges can have a significant impact once you’re talking about the information shared by the populations of entire countries.
Now some companies publicly release the weights for some or all of their models, and this openness can be helpful. If we know that a model has been directed to skew its outputs in certain ways, we can share those skews with others and hopefully inoculate people to that bias. And similarly, when those biases have to be put on display there is an incentive to not put one’s thumb on the scale. Although as seen from the Grok examples the fact that there are problems with a model does not stop people from using it.
But a similar problem exists from how those models are trained. Because these models work by collecting massive amounts of data, often scraped from the internet. And the data itself can carry a variety of biases, both in terms of the data itself and what is or is not used (or available to be used). Fundamentally, that training data is going to be built out of the social and political structure of the system that creates the underlying data. This means that any AI can only ever be a mirror of the world that created it. If you think that perhaps some facet of the world is broken, AI cannot help to fix it.
So more people becoming reliant upon AI for information turns into a homogenization of society. And again, this need not be a directed result. You can have those with the keys to the system pushing buttons in ways to push people in a particular direction, but you don’t need that kind of button-pushing. Just the fact that the AI serves as this mirror means that people will produce so many reflections of what already exists.
And we should not ignore the way in which art is fundamentally denigrated through all of these processes. These models being trained on existing material means that the creation of new prose or art requires drawing information from existing artists. When you ask ChatGPT to write a poem for you or throw a prompt into Midjourney to create a picture, those models are using data obtained by scanning books and poems and art from as many places as possible, and then putting together words or pixels in such a way that it sufficiently imitates the original materials. These companies tend to not care much how their training data is gathered, as evidenced by so many stories of websites implementing various “locks” that are meant to prevent AI scraping that are then bypassed. Everything – this blog post, a drawing you share with a group of online friends, or anything else shared on the internet – is seen as fodder for AI models.
Just as the input of art is treated as so much data to be consumed, so is the output of art treated as a mere collection of lines or letters. As long as the product looks or sounds neat, there is no need to think further about it. I discussed before how AI models don’t fundamentally “work” when it comes to producing something artistic – it produces things that look like art, but lack the characteristics of actual art. The choices for how a piece of art is composed or a piece of writing is structured, the style employed, the techniques used, and all the components of the effort that underlies that artistic endeavor is demeaned in favor of a mere end product, a product which can be produced quickly and cheaply. Art has faced these problems well before the advent of AI, and yet AI exacerbates these problems as the people behind that art can be cut out of the process entirely.
This denigration doesn’t mean that true art disappears from the world. But it does mean that the avenues for people to practice and make a living from art – and likely with that the methods of teaching art – start to shrink. People still draw and write because they enjoy the process and feel good having created something. But more and more have to do so for nothing more than that love.
The treatment of information and art as mere parcels of data to be generated at the whim of the developers of an AI or its users further cements the idea that everything is a commodity to be bought and sold.
AI, Video Games, and Game Development
This blog being ostensibly focused on video games, I want to pull all of this together into how it impacts video games and game development.
Game development itself straddles this kind of middle ground between the artistic and the technical. Games are built off of code, which might make them seem like they fit within the realm of AI usage where the models are (potentially) functional in a meaningful sense. But games also involve a huge number of artistic components – the worlds they might be set in, the environments that specific parts of the games take place in, the stories that motivate them, the design of the characters that inhabit the world, the dialogue of those characters, and so on. And even if we assumed development could be cleanly divided between artistic and technical tasks, the push to adopt AI is not constrained to only the latter. AI is being pushed into every corner possible – AI being used to generate concept art, to design environments, to write dialogue, etc.
To my knowledge, there is a significant push from many rank-and-file developers against the use of generative AI. Various programs to procedurally generate content within a game isn’t alien to the process (the genre of roguelikes being one of the most prominent examples of this), and yet the adoption of AI is recognized as a step too far that can harm those developers in the long run. Some studios have used AI in some parts of their process. Clair Obscur developer Sandfall Interactive received a lot of flak when it was revealed that some content – most notably some posters found on various poles in parts of the game – was generated by AI, allegedly as placeholder content that wasn’t replaced later in the process. Even that purportedly small amount was sufficient for Clair Obscur to have an award stripped from it. And despite the effort that undoubtedly went into the game by the developers, that call was rightly made.
Because the usage of AI poisons the well. If AI is used to generate art at one stage of development – especially when that usage is kept secret – then there’s no way to be sure that it isn’t used at other stages. If AI is used for concept art, why would anyone believe it wasn’t used for the characters? If it’s used for environmental dialogue, why would anyone believe it wasn’t used for other parts of the writing? One of the core defenses when these cases arise is that AI is only used in specific parts. Perhaps AI is used to create an original draft of something that is then revised by a human. But these defenses fall flat. Because there’s no way to demonstrate the claim – it must be taken on trust. And where the mere use of AI is seen as a violation of trust, where the audience trusts that the developers are creating something artistic with an intent, there is no way to earn it back by defending the original violation as minor.
And again, to my understanding where this usage happens among existing studios it is largely a top-down imposition. There are a lot of game devs on social media who are happy to talk about how much they hate generative AI. Whereas a lot of the praise about AI from the game industry is sourced to the heads of studios. And perhaps in part those rank-and-file employees see the problems I’ve already described. In an occupation that is extremely volatile, where jobs can easily be cut and studios closed because some publisher wanted to downsize for a temporary boost revenue for shareholder reports, where poor management can lead to psyche-breaking levels of crunch, AI poses a distinct threat. Just like in other areas, there’s little reason to believe that the widespread adoption of AI will result in better working conditions or better pay or job stability. If you can generate concept art and then have someone modify it a bit, why have multiple concept artists? If you can generate tons of dialogue, why have multiple writers? If you can have those lines voiced by an AI, why hire voice actors?
But outside of the standard process of game development, AI has resulted in numerous other issues. These AI tools are sold as something that can make pretty much anything from the ground up, and you as a user just need to write a prompt. And so you can make a game with absolutely nothing. Just ask the AI to write up the code to accomplish whatever you need, and generate the art, and write the dialogue, and so on. And this all results in a proliferation of AI slop. “Slop” here refers to two things. First is a product that is made thoughtlessly, usually to turn around to make some quick money or share for clout. Steam and similar online platforms have seen a rise in clones, where a popular game is copied numerous times in the hopes of not just cashing in on a trend but actively confusing potential buyers. Second is a product that is made with an intent, but lacks an understanding of what the user is trying to create. Someone who genuinely wants to make a game but lacks the artistic skills or technical knowledge might turn to AI, but in doing so produces a game that is bland – they may be sincere, but because they have not put in the effort their game cannot appeal to others. Now and then you can get a game that people enjoy playing, but these are going to be few and far between. And even where that was true before, the proliferation of these tools is the difference between pushing new games through a firehose and an actual flood.
In line with that first meaning of AI slop is the way in which these tools also exacerbate problems of trend chasing. A major pitfall of the game industry writ large is how often we see publishers/producers push studios to make games or incorporate mechanics to mimic something that is currently popular. Probably the biggest overall trend are service games: online games that keep players engaged with regular content updates and numerous microtransactions. Stuff like Overwatch or Destiny and so on. Now if a new game drops that amazes everyone, it can take three to five years for a studio to work fast on a game from scratch to try and ride that same wave. But if AI can be used as a shortcut for so many of these processes, then that time can be pared down. Less and less scrupulous studios/publishers can use these tools to cash in on a trend, which in turn pushes those higher up the chain to do this more and more. The short-term gains of adopting AI and chasing the next big thing establish a feedback loop that is unhealthy for the industry as a whole.
And I want to take a moment to talk about the whole decompilation/recompliation[1] project. For years there have been efforts to essentially break down and rebuild games as a way of preserving their code. There is an impact where successfully decompiling a game can allow it to be shared more widely, but usually these projects do not exist for the purpose of making those games available – emulation already exists for many of the games in question. AI has recently been applied to this process, resulting in all sorts of games being recompiled and shared, often hosted on the internet such that you can play them on your web browser. That wide accessibility to these games at once seems cool: all sorts of people can now play a game incredibly easily. No need to seek out and download emulators and ROMs or ISOs.
But since the purpose of these projects has primarily been preservation, the use of AI has overshadowed these efforts. Indeed, a major decompilation community has effectively disbanded because other users have jumped in and had an AI “complete” the project. They might take whatever a community has already decompiled and then tell the AI to finish what’s left, or just start from scratch. In either case, these AI then just take the hard work already done by others to effectively put a bow on it and repackage it as their own effort. It is, in effect, a way to just stand up and claim credit – to receive applause for breaking down a game and making it available for others. Which takes us back to the problem of the denigration of art. Just as the use of AI to generate art by taking from thousands of artists turns the concept of art into just a set of lines and colors that look nice, so does the use of AI to decompile a game morph decompilation from a data preservation effort into an effort to simply produce decompiled games. People can work for years to decompile Pokemon Heart Gold/Soul Silver and publicly document their progress, and then someone else can come along and spend a bunch of tokens to ask an AI to do the same thing, which then just copies the work already done. This explosion of recompilations needs to be understood within this context, and the impact it has on a community of people engaged in sincere work.
Everything here is a showcase of the underlying problems with taking the narrow view of AI as a mere tool. The first thought about all of this is that surely the use of AI is good. It produces something cool (“hey, here’s a game you can now play for free on your computer”), it does it in a way that feels neat (“wow, this would have taken years, but I did it in just a couple months with AI”), and in some cases it can feel like a way to grab control back from corporations (“these companies want to sell me a game for 80 bucks, but now I can play it for free”). The aesthetics surrounding AI suggests that it is good. But underneath everything are incredibly harmful impacts. These impacts are not on display in the way a new game or a piece of art is, nor do they take place immediately. And so by focusing solely on the product, the debate centers around AI as a tool and whether it’s a good tool or a bad tool. But AI is not analogous to a hammer. A hammer cannot disrupt the livelihoods of an industry. A hammer cannot reorient our approach to art and creativity. So to talk about AI as a tool is to rip it out of this context and ignore its effects on game development and the people within it in order to focus on how you get something that looks cool.
Concluding Remarks
This essay is woefully inadequate.
Any of the topics I’ve discussed above can be explored in far more detail. And on top of that are so many angles that are important to investigate. There’s the construction of the data centers that can be disruptive to nearby communities. There’s the water usage that goes into training these models. There’s the energy drain that comes from maintaining the data centers, which increases the costs to everyday people to simply live in their homes. There’s the drain on computing resources (most notably RAM right now) that makes access to computers more expensive. There’s the case of these AI companies buying, scanning, and destroying rare books, treating knowledge as a commodity to be hoarded. There’s the threat of AI psychosis. There’s the problem of deskilling as more intellectual activity is offloaded onto AI.
There is so much more that could be covered, but my hope is that what I have covered here is sufficient to support my core claim:
Using AI – this generative AI that is currently being pushed onto us – is unethical.
There are different layers of this claim. At the top are those responsible for creating and perpetuating these AI models. The companies that develop the largest models and the companies within their orbit supporting those models by providing infrastructure. These companies are engaged in a sort of arms race to develop models that outpace their competitors in the hope of producing what is referred to as Artificial General Intelligence (the thing we actually think of from the term “AI,” a true intelligence and not just a machine that puts words or pixels together in a way that mimics human output). The goals of these companies is claimed to be to create AI for the good of humanity, and the heads of those companies will publicly state that they share concerns about various problems and bad usage. But these same companies don’t take effective means to reign in any problems, continually sell the models as extremely powerful and able to serve as effective replacements for various occupations (including and especially art), and hype up any new releases as potentially dangerous if it should fall into the wrong hands. Reportedly many of the people working in this field believe that AI may result in something catastrophic for humanity as a whole. Ignoring whether such predictions are true (I tend to believe they are exaggerated), to continue such work under this belief is monstrous. And such rhetoric is often shared openly by the leaders of these companies. So either these people are lying, or they are actively building a technology they know to be extremely dangerous. Either outcome speaks poorly on them.
The next level down are those pushing this, particularly managers and any other prominent persons and companies that are extolling the adoption of AI. These people receive less condemnation because they are neither directly building the tools, nor do they perceive the tools as dangerous. Yet they are still pushing AI as a form of exploitation. Going back to the efficiency topic from before, managers and other influencers are not promoting AI as a way for people to become more efficient and thus earn more money or get back time for their lives. They are pushing this as a way to either get more out of the workers they have, or reduce the headcount of their companies (and perhaps both). There is little to no concern for the persons who are fired, nor is there any question about whether asking people to churn out more work for the same pay simply because they can is fair. Because such people don’t care about these questions. And so this exploitation must earn a substantial amount of ire.
The lowest level are those everyday people who make use of AI. And within this category there is still variation. Those who voluntarily use these tools to generate artwork or code – even where it is good or at least serviceable – are helping to perpetuate these systems and suggest that they are good to use. Those who are compelled to use them (for example, a programmer who is required to use AI or else be fired, even if they do not find such tools useful) are still aiding in that perpetuation, but are coerced to do so. Each individual usage is going to seem small. If I generate an image or ask ChatGPT to draft an essay or make a game with some AI art, those uses can feel innocent. I could, for example, try to make a game. I’m not good at producing art, so I could have an AI model produce various characters and environments for me. And not having the knowledge of coding, I could ask Claude or some similar model to write the code and debug the game and so on. And I don’t have the funds to hire a voice actor and can’t really do all the voices myself, so I have another model produce the voice lines. All of this might cost me a decent amount of money, but far less than hiring an artist and a programmer and some voice actors, and might thus be within my budget. A project that would cost $100,000-200,000 instead costs maybe a tenth of that. And so this all feels okay. But as those choices get universalized, we see a larger impact. Various artists and programmers and other personnel all get sidelined in favor of using AI. The people behind everything get thrown away.
And that’s what I am hoping to convey through all of this. No use of AI is truly isolated. Every question posed, every image generated, is all promoting these models as legitimate tools for the work it is performing. And it all contributes to the exploitation and theft and other harms that come from the development and deployment of these models. There is no escaping this association.
Now it would be simple enough to say that the companies behind all this should just stop and the problem goes away. But they do not stop out of good will. If that were enough, they would have already done so. The companies stop when there is enough pushback to make these models unprofitable. When people decide that they do not want to use these models anymore. Hence why it is important for everyday people to do what they can to divest themselves of AI. As these tools suffuse more of our lives that may be easier said than done. Even if we can’t remove AI 100%, we should still strive to remove it as much as possible. Just as each use helps sell the idea that AI is valuable, each removal helps showcase that we don’t want it. Part of how AI continues to integrate into technology is by people allowing it, by seeing it as inevitable. And likewise by seeing any small failure as reason to give up entirely. There will be substantial rhetoric around this issue that in order to oppose AI one must oppose it in any and all forms perfectly. That any usage, known or unknown, willing or unwilling, speaks to a moral failure. But the purpose of trying to push back is to take the stance and work toward that goal. To promote human-made art. To avoid AI-made products where possible. To stake a position that signals that this current version of AI is not acceptable.
To end this, I want to provide some basic rules by which we can understand when AI and its use becomes ethical. Even the current models could be ethical, but to do so they must be built and deployed in a way that actually supports people as a whole, and not simply AI companies and managers. These rules are not exhaustive, but rather some general principles to begin.
- Construction – Current AI models are built on a large amount of theft. Artistic works such as books, poems, films, drawings, and all such other things are scraped from all possible sources to serve as training data. This is all done with no concern for whether a given individual consents to their words or artwork being used for this purpose. To be ethical the training of an AI needs to comport with ethical rules which consider where that data originates from.
- Utilization – As I’ve gone over before, AI is sold as something that can effectively accomplish a multitude of complex tasks, often with little or even no oversight. Many users may be aware that oversight is needed, but the vast majority of users are not going to understand this because that’s not how such models are being presented. In order for the usage of AI to be ethical, it needs to be perceived as a tool which is limited, and where users are fully aware of those limitations. If an AI is only suited to specific, technical tasks which requires an expert to review the output for accuracy, then it should be marketed and deployed that way. Such deployment may mean it turns into a tool applied only occasionally, and that’s fine. Because seeing it as a general tool that can be applied anywhere is what results in the belief that the people doing the actual work are expendable.
- Humanity – Usually when we hear “the good of humanity” talked about in the context of AI, it’s as a generic good value. But AI working for the good of humanity requires specific policies and actions that are not upheld at present. In the event that any AI could be so reliably productive that substantially less human effort is needed, then society as a whole needs to be reoriented around that decrease in effort. And the overall social discourse needs to reflect that reorientation. If less work is needed, then people should have their time returned (without losing their pay). People should not be losing their jobs so that they can be replaced by an AI. And we should be seeing a Universal Basic Income as a necessary component for sustaining people. If you do not see these or similar ideas being proposed and fought for and implemented, then AI has no place within that society until it is ready.
And we should not ignore the potential mistreatment of AI itself. Right now the models that exist are not “true” artificial intelligences, and as such there is no mistreatment. But the ideal of AI and what these companies are building toward is a true intelligence, which will undoubtedly involve some degree of sentience. If we were to build such an AI for the purpose of being subservient, we would effectively be creating a being in order to enslave it. So while we do not need to contend with this problem right now, the ethical use of AI must be aware of and ready for this problem of exploitation of other beings.
And so I will end with the reminder that as neat as the products are, we should not be taken in by the aesthetics of AI.
[1] I’m going to use these two terms interchangeably, even though they are different tasks with different goals. This choice is that while hobbyists in these spaces are aware of what they’re doing, the way in which AI is applied to these projects is fundamentally similar, and so poses the same basic problems.
