A Critical Approach To AI In Digital Art

In the courtyard of a gallery in Berkeley, a young artist on a panel is answering a question from a curator about what art could emerge from a future that seems, at least in this moment, to be dominated by the intertwined spectres of climate disaster and resurgent American fascism. The show has been organized to showcase works that engage with generative AI, as a kind of numerical manifestation of the Lacanian real, where indescribable amalgamations of text and image manifest into the symbolic order when conjured by prompts (or so the curators’ statement implies).

After a series of tepid responses from the older AI-using artists gesturing to their general unease about the world they’re bequeathing to their children, the young artist begins with a parenthetical: “aside from the mass extinction event,” she says, this crisis is an opportunity for new forms to emerge.

In the years since this exhibition, her trivial dismissal of ecological disaster has haunted me, as more and more digital artists I know turn to using generative AI. The majority are people using coding LLMs to create software tools that they either couldn’t complete on their own, or didn’t have the time to. Some are centering the image and video products of generative AI in their work. As an artist whose day job is in software engineering, I can’t pretend I didn’t see this coming; between bosses embracing the tools and their improvement in the span of the last year to be genuinely helpful in many cases, it became clear to me recently that, at least when it comes to coding LLMs, generative AI isn’t a fad.

Image Generation As Aesthetic Black Hole

I want to begin with a demystification: “AI” is just a marketing term that captures a number of distinct but adjacent technologies under its banner. Large Language Models are distinct from image generation models, which are themselves distinct from image recognition models. Because this is an essay about art, I suspect readers are expecting a focus on image and video generation models rather than the creation of software, so let’s get this dealt with off the bat: those models are an abhorrent scourge that I don’t think any artist can engage with in good faith.

They represent the aesthetic sensibilities of fascism, and the way the models are constituted makes them incapable of producing other aesthetics: the blurry, waxy faces, the too-perfect lighting, everything rendered in high dynamic range–this is a look that, over these models’ existence, has become inseparable from the far-right. To the average queer or non-white person in America and western Europe, these images look like all of the things that have made our lives miserable over the last few years. The Trump administration is rife with low-effort, prompt generated AI art courtesy of tools made by Sam Altman, Mark Zuckerberg, and the like. We can, of course, return to Walter Benjamin, who wrote about reactionary aesthetics in his own time by saying “the logical result of Fascism is the introduction of aesthetics into political life.” [5] AI tools provide an unending source of material to aestheticize politics with ease; every meme creates cues that form a visual reality for reactionaries to live inside, an aesthetics of exaggerated crisis that demands a strong hand to put it to rights.

Once the political has been reduced to aesthetics in this way, policymaking becomes a question of reinforcing these aesthetics. National Guard troops are sent to Chicago because the aesthetics of fascism demand riots that can be quelled. A liberal might believe they would be left alone if no riot forms; to the fascist, a riot can be conjured from a crowd just as easily as ChatGPT can be prompted. The pleasure of violent repression calls out from the image generator, and the world must then be made to conform.

This is why “aside from the mass extinction event” comes across as a particularly honest expression of such a political project. Benjamin claims that fascism’s “self-alienation has reached such a degree that it can experience its own destruction as an aesthetic pleasure,” and so this artist looks to the pleasure of aesthetics in a society still trying to reckon with the destruction caused by the political movement that it evokes. Embracing this nihilism, whether intentionally or through exhaustion with other forms of aesthetics, is embracing a world in which capitalism’s death spirals only accelerate, and bury us all with it.

There are, however, possibilities for life. We may yet make it to a better world intact. To do so requires an effort, as Benjamin puts it, towards “politicizing art,” to look at machine learning not as spooky tangles of inscrutable math, but as concrete things imbued with the character of the class politics that created them. To reckon with systems, rather than magic.

So, what are these systems, and what do they do? Most of them are prompted by text: you take a sentence that asks for something, and the machine generates an image (or piece of code, or text) that corresponds to what you ask for. It does this by being trained on existing datasets which comprise billions of images, books, text files, and codebases to provide an amalgamation of these appropriate enough to answer your question (obviously it is vastly more complicated than that, but this is the gist). In Capital Vol. 1, Marx talks about how capitalism takes historically disparate methods of production and concentrates them: “the transformation of the instruments of labour into instruments of labour only usable in common, the economising of all means of production by their use as means of production of combined, socialised labour,” in order to be exploited by the capitalist. The conditions of production under capitalism are social (we all broadly work together and build off of what others have done) while the profits are kept individual, for the owners of the means of production.

Machine learning systems, then, seek to take existing disparate forms of labor and turn them into something a capitalist can profit from. Marx, in his time, saw this same process taking place as local artisans were supplanted by factories with hundreds of those former artisans producing goods far cheaper than anyone could before. We could call that soulless (and certainly, factory production is soulless). But for goods like boots and coats, they have largely utilitarian purposes, so a soul is perhaps not a strict requirement of the outputs of production. I mostly need a coat to keep me warm. But when it comes to art and art objects, the aesthetics of the produced item is the entirety of its value, and the object is judged almost entirely on its aesthetic principles, which are socially determined. The aesthetics of the production process are very clearly visible within the aesthetics of the produced object; AI generated images may be selected by an artist, but because they are created via an amalgamation of past images directly, they cannot escape immediate identification with their AI production process. This, then, is the crux of the problem with AI generated images: the soullessness with which they are produced necessarily creates an aesthetic that evokes its soullessness in how it’s perceived. It’s the apogee of capitalist aesthetics; no wonder it’s been adopted by fascists.

Machine-Written Code

When the product of these systems, then, is not strictly an aesthetic object, but a tool used in the production of other things (aesthetic or otherwise), we have to talk through problems of externalities, rather than immediately apparent aesthetics. The aesthetics of Adobe Photoshop may bleed into photography, but they don’t define it; it is the individual or collective human intentionality of photographers or digital collage artists that comes back into the fold here, not just a vulgar machine, as is the case of AI image generators. While code can itself be aesthetic, it is rarely the art object that’s being produced, and more often a byproduct of its production. Of course, counterexamples exist (consider: algoraves, etc); in those cases, human intentionality remains paramount, and, as in image generation, I am far less interested in seeing the outputs of the LLM rendered as an aesthetic object. But, in the vast majority of cases, code is a transparent tool that enables art making. I rely on people to make that distinction themselves and live with the resulting ambiguity.

Let’s turn our attention, then, away from image generation and towards coding LLMs, tools which rely on vast corpuses of code to produce tools, which are then used to produce aesthetic outputs via human intention.

It’s not hard to see why artists may be interested in using a tool like this: not every artist is a software engineer, and even the ones who are have limited understanding of certain types of software. Being able to create tools that allow people not trained in, say, embedded systems or mobile development to display (human generated, mind you!) images on microcontroller-driven screens or capture images in a novel way through a camera phone app lets more people create art that relies on new, computer-centric forms. It allows for fine-tuned, intentional control over the aesthetics of the produced art object that was previously contingent on expert knowledge. Going back to the Photoshop metaphor: in the early 1980s, the only people who could create photo manipulations and digital collage were people with years of darkroom experience. By the late 1990s, anyone with a PC could do it at home (without the chemicals). I expect coding LLMs will have the same impact on digital art, and I can see why artists would want a machine that produces these outputs. You could suggest an artist hire a software developer for this, but this is not a serious argument. Paying someone $200 an hour for a minimum of 40 hours of work to write a one-off application is not within the budget of nearly any digital artist except for those at the top of their field. Plus, the flexibility of being able to modify it for new scenarios as a work evolves its presentation is critical.

Furthermore, I would argue that making the creation of software more accessible is a net good. Software engineering as a field is overwhelmingly male and white; the knowledge of how to write software, then, has a deep race and gender bias. My sincerely held belief even as someone who possesses that knowledge, is that we all benefit socially by allowing people of different backgrounds to create art that examines how we all interact with digital technology. While the practice of software engineering becomes more specialized and, as a result of LLM adoption, shifts away from “writing code” and towards “the maintenance of large systems,” there is room for artists who don’t have a software background to produce basic code via LLMs without the skills necessary to maintain it. Thankfully, maintainability is genuinely less important when we’re talking about tools used to facilitate small artistic processes. We don’t demand artists learn everything about the intricacies of how a compiler takes Processing code and turns it into JVM bytecode, for example, and high levels of abstraction are more tolerable when the goal is to create a work and not maintain a system for others. There is still the opportunity for craft here as well; while LLM-related cognitive impacts are real and have been well studied, there’s also the potential for a curious artist to explore the space they’re suddenly capable of working in. Given intentionality, LLM codegen can open the door to learning the craft of a field that was previously considered inaccessible.

But LLMs have specific valences in the western context, which should be more than clear now as 2026 draws to a close. The potential for overwhelming energy and water use (the ecological catastrophe, discussed in the introduction), poisoning the air of Black communities in the American south, and of course, the American AI companies’ themselves having embraced a far-right authoritarian and imperialist project. It’s understood that every tool depends on the labor of others (see Marx, above), often in less than ideal circumstances, and every capitalist is an enemy of the working class. But to return to a tired metaphor: every one of us is carrying a smartphone in our pockets that was produced under labor conditions that people in America would find abhorrent, likely made by a company that has cozied up to Donald Trump. This sentence, almost verbatim, is used as a rhetorical cudgel to either get people to abandon their technology and live as a hermit in the forest (unlikely, though some have tried) or to get people to stop talking about the labor conditions involved in the creation of an iPhone. I’d like to suggest a different approach.

As artists, I understand our practice to be an interrogation of the things others take for granted; aspects of identity, surveillance, and, yes, conditions of production, are up for examination. With that in mind, we cannot simply gloss over any aspect of this as an expedient. If there are alternatives to ecocide, we have to investigate them. I would not hold any artist to the standards of our iPhone-less ascetic, but I would expect us to at least try to make ourselves aware of the least-offensive way to achieve our ends. Thankfully for us and for the planet, there are less harmful options, but you won’t find them in the United States.

Minimizing Harm

Before talking through the ecology and politics of generating code via LLMs, it’s important to understand that a large model works in two stages. The first is training. A model needs to ingest millions upon millions of text files (code and otherwise) to understand how to respond to potential queries. Once that’s been done (in a process that takes years and millions of compute hours on high-end GPUs), you can then prompt the model. This second process is called inference, as in, the model “infers” how to answer a question. Each individual inference query is cheaper and more efficient than training by orders of magnitude, but there are substantially more of them over the lifetime of the model. You, as the end user, never do anything that directly impacts training; every time you interact with an AI model, it’s doing inference.

The most popular non-American models are what’s called “open-weights”; this means the results of the model’s training are made public, so anyone can run inference on their own computer. There are two major benefits to this: if you’re paying a cloud provider for inference, you can choose a provider with the least ecological impact. If you have a computer capable of running these models at your house (a mid-tier gaming machine is fine), then you can run it entirely on whatever electricity you have at home (generally, far more sustainably than Anthropic’s Claude running at Elon Musk’s Colossus datacenter). These models are quite capable; for the small projects that artists are most likely to engage in, they’re more than powerful enough, and it’s possible to run these on entirely renewable setups. My house, for example, is powered by solar panels, and the water cooling in my PC is a closed loop; I’m able to generate code with essentially zero marginal ecological impact (I suspect I’ll write a follow up on exactly how to do that, but this one is already a few thousand words, so it’ll have to wait).

A thornier subject, then, is resource use in training. I’d argue that, for open-weight models, this is less important than it may sound. Using an open model on your own PC gives zero benefit to the company producing that model. There is no additional incentive for them to use more energy to train more models if you just run their model entirely on your own infrastructure, paying for your own compute, instead of paying them for cloud usage. Qwen does not see any revenue from me or any other artist using their model on my PC. Contrast that with using OpenAI, where they gain revenue from each ChatGPT query you make.

If you don’t have a spare gaming computer, there are options for cloud-based inference. While Anthropic is buying up even more fossil fuel generators that poison the air, Mistral, a European company, has built the data centers required to train its newer models in France, a country whose grid is powered by 95% non-carbon-producing energy. Mistral Large 4 is a perfectly capable model, equivalent to mid-tier American models, and the data centers used to host and train it run on a mix of largely nuclear, solar, and wind power. We don’t have specific data on water use for this generation of models, but Mistral Large 2, a previous generation model, over the entire lifespan of the model (ie, training and 18 months of prompting) used half the amount of water that GPT-4 is estimated to have used for training alone. In China, which generates about 40% of its energy from non-carbon-producing sources, the state is using industrial policy to incentivize building datacenters in places with large amounts of renewable energy. The US, which produces a similar share of energy from fossil fuel, is doing the opposite with its incentive structures.

And now, a return to political economics. In the United States, AI companies are selling their products to capitalists as a way to replace workers; in China, the state has taken an active role in mitigating job loss, and courts have ruled that workers cannot be fired if replaced by AI. The European Union has similar laws in effect governing mass layoffs as a result of automation. The economic incentive structures around AI in non-American contexts are also completely different: the US has priority access to the hardware needed to train AI models, and an economy that increasingly is reliant on this industry. They are investing heavily in the most capable models at any price–ecological, political, or human–because their sky high valuations demand it. In order to pay back creditors, the AI built in America has to be able to disrupt every possible industry (something which, despite its ability to write software, I still believe is more or less impossible). Meanwhile, Chinese state policy is to find useful areas where AI can be beneficial in the here and now rather than to focus all effort on building some massive general-purpose AI model. They also have the pressure of American export restrictions limiting their access to the fastest GPUs, so they necessarily have to be leaner than their American counterparts; the models they produce are then far less resource intensive.

This is not an argument for anyone with no use for AI to use it, but rather an acknowledgment that artists are already using these tools, and they’re using the worst possible versions of them. Every time I speak to people about this, it’s always explained to me in shameful terms: “well, I know it’s bad, but I used Claude for this software” or “I was really in a rush, so I wrote a script with ChatGPT.” To my mind, this is not an acceptable way to engage with this technology when alternatives exist. But, by and large, we are not talking about alternatives or explaining them. The only people who are aware that models outside of the big American labs exist are techies or people who run local AI services as a hobby (and mostly seem to generate porn with them). The popular discourse around the subject has settled onto a reflexive model of abstinence only, with severe rhetorical penalties for those who step out of line.

This is not conducive to creating a world where people can examine the impacts of what’s becoming a very useful but compromised technology; it’s creating a world where American artists use American AI that is actively killing Iranians and Black Americans and simply just don’t tell anyone they’re doing it. If there were smartphones that mitigated the labor conditions and conflict minerals involved in their production by 10%, we’d be commending people for buying them. Instead, when someone talks about using more ethical AI models, they’re often pedantically told they don’t actually need to use AI and that they’re just being lazy. So, they hide it; they don’t write about it in their artist statements; they don’t disclose its use. And as coding LLMs get more and more useful, more people will use them. Without a structural change in how we approach the subject, I expect they’ll just lie about it, and keep using the models made by companies with blood on their hands.

This is the first blog post in a series of at least two where I attempt to lay out a vision for more-ethical use of coding LLMs in the production of digital art. The next will cover running your own models locally, and a comparison of cloud providers’ environmental records for those who can’t. Stay tuned.