Recently, I fulfilled a childhood dream. My debut fantasy novel sold to a publisher; my words will soon be packaged behind a pretty cover and shared with the world. Along with this victory came an urge to look back on the path that led me here—a path I can more-or-less trace through Twitter, ever since I joined the creative community there a few years ago.

So I went trawling through my tweet history, smiling at old pitches and snippets, riding a wave of happy nostalgia. Until I came across a post from late 2021 that stopped me cold.

It was a graphic of a psychedelic-looking motorcycle on a wasteland background, captioned with a breathless declaration: AI art is awesome!

The post was part of a trend going around the Twitter writing community: using an AI art generator to create fanart for your book. This was before anybody in my feed started talking about the dark side of AI, before I heard critics point out that AI generators are trained on stolen artwork, before it occurred to me that synthetic art could be theft.

In 2023, I read that tweet and felt a wave of shame so strong I reflexively hit Delete.

I have long felt like an imposter in both tech and creative circles, straddling an apparent divide between the two and feeling amateurish and outclassed by anybody who specializes in just one. When I was studying computer science at Stanford, I attempted NaNoWriMo twice but failed, overwhelmed by coursework. Tech was literally getting in the way of my creativity.

These days, I oscillate between worrying that writing silly stories about magic and monsters is a distraction from the real-world opportunities of the software industry, and worrying that working at a FAANG makes me a corporate sellout and a colonizer.

AI’s first steps into “creativity” felt like the long-awaited merging of my two worlds. I was exhilarated; I wasn’t thinking about ethics. When people started pointing out their concerns, my first reaction was dismissal and a tiny bit of frustration. I had been encountering people with seemingly baseless fears about AI for a long time, and I thought they were just wary because they didn’t understand how the models worked. They thought AI was dangerously close to becoming sentient, and I was supremely confident that it was not.

Reader, I was missing the point. Badly.

I’m still not quite sure how the average person without a CS degree conceives of something like a neural network—whether it’s necessary to understand matrix multiplication and linear optimization in order to grok that machine sentience isn’t anywhere close to reality. But I do know that AI doesn’t have to be sentient to cause harm.

If you’re part of the publishing community, you know what happened next. The buck didn’t stop at visual artists. OpenAI released ChatGPT and—almost immediately—people realized with the right prompts, they could make the chatbot write fiction. Spammers started churning out AI-written stories and clobbering magazine slush piles, Clarkesworld put out a statement and closed submissions, and my group chats blew up with horrified writer friends wondering if this spelled doom for their careers.

And I realized in dismay that my two worlds weren’t so much becoming one as one was eating the other. AI’s impact on the small, tight-knit world of speculative fiction was immediately and unequivocally negative. There was no upside to what was happening, none of the brilliant meshing of technology and art that I had envisioned in the beginning.

This was a disaster.

For brevity’s sake, I’m going to call both AI-generated visuals and AI-generated prose “AI art,” making distinctions where appropriate. While there are disparities in how AI has impacted subcommunities within the larger creative community, writers and visual artists are more alike than not. AI-generated visuals might blend in better than AI-generated prose for the moment, but it’s not wise to assume this will always be the case. We’re all on the same side here.

ChatGPT and Midjourney are both examples of AI, but they use different underlying models (and obviously are trained on different datasets: text versus images). ChatGPT is a Large Language Model, or LLM, which—in simple terms—takes a text input and predicts what will come next, using a statistical model.

Midjourney is a little harder to explain. It relies on a Generative Adversarial Network, or GAN, which actually consists of two dueling neural networks (if “neural network” is a fuzzy term to you, feel free to think of it as a “model” or just “an AI”). One of these models is a generator: it produces a fake image. The other model is a discriminator: it compares the fake image to a real image from a training set and calculates how similar they are. The discriminator feeds the result of the calculation to the generator, which updates its model so the next image it produces is a little more convincing, a little closer to the real thing.

It’s like an art student who has never seen a lion learning to draw a lion over repeated attempts, while their teacher—who moonlights as a zookeeper and has seen plenty of lions—grades each attempt for accuracy and suggests what they could do better.

The product called Midjourney that people use to create art is actually just the generator. Once the generator is fully trained, you don’t need the discriminator anymore. The reason I’ve gone to the trouble of explaining all of this is because some AI defenders claim Midjourney doesn’t use any references: it “creates images from scratch, without using any training data,” conveniently subverting the common criticism that AI art generators are trained on stolen work. This argument is partially correct, but misleading: the generator didn’t see any training data, but the discriminator did, and it used that data to train the generator by proxy.

Another popular sentiment thrown around in defense of AI art is what I like to call the Photoshop argument. The Photoshop argument states that AI is just a tool, like Photoshop, and so you can’t assign ownership of the generated output to the AI, but to the person who fed it the prompt. This argument is used to claim that a person using AI is no different from a person using a paintbrush or a tablet, and in all cases that person is an Artist and should receive the respect and privileges afforded by society to those with that title (ha).

In reality, the input prompt given to an AI is more like a search term you put into Google, if the results were randomized, combined, and run through a filter to make them look different from the initial set. You wouldn’t call the Google search engine a “tool” you’re using to “create” the results, would you? It’s definitely not the same as going out, doing the research, and writing a result for your search term yourself.

By and large, the people wielding AI like a sledgehammer in creative spaces are not AI programmers. They didn’t take computer ethics classes in school or learn how to build and train a neural network from scratch. I’m saying this not to shift the blame from programmers to anonymous scam artists, but to make the landscape clear: everybody has access to AI now—in its current form, with its current capabilities—and nobody is going to put it back in the bottle.

I’m not going to give much time to the scammers here. Nothing I say is going to convince them to stop. As long as there’s even a hint of duplicitous profit to be had, they’re going to do what scammers do. AI is just the latest tool in the arsenal that spam callers and robotexters have been building for decades.

But I do want to address the people who think bringing AI into creative spaces is a morally neutral or even positive thing to do. Because they exist; these people are my colleagues, people with whom I share an alma mater or a current or former employer. Their arguments are numerous and some of them are even persuasive:

“AI can make art more accessible to people with disabilities.”

“AI is just a tool, the same way a digital brush or a filter is a tool.”

“Artistic meaning is in the eye of the beholder; it doesn’t matter who or what made the art, or if they were capable of expressing original thought.”

“AI is changing every other industry—get used to it.”

“If AI can tell a story as well as a human, then it deserves a shot.”

To all this I say: okay, but why?

I understand the appeal of playing with AI, testing its capabilities. Curiosity is a basic human impulse, the impulse that drives scientific advancement and human progress. Who is more curious than a writer exploring entire worlds that only exist in their mind? Curiosity is the most significant trait that creatives, scientists, and engineers all have in common. But what we’ve seen so far from people using Midjourney to sell picture books or ChatGPT to clog magazine submission queues goes beyond curiosity.

It’s something else, and I’ve been trying to figure out what it is.

AI art has succeeded in one thing: pissing off a lot of people in a short span of time. Understandably, most creatives would probably support a total ban on AI art in their communities. It’s an instinctive “get this shit out of here” reaction. Totally normal in response to something that’s actively causing harm!

But I don’t think AI art is going anywhere. So is there any way to ethically use AI to create art? Is there a place for it in the creative process?

For me, it’s a personal question: can my two worlds coexist in peace?

Asking questions is a good place to start. It’s tempting to cling to people and arguments that make you feel good and safe, rejecting anything that feels uncomfy. In the AI art debate, as in every other debate, you should embrace being uncomfortable. That doesn’t mean you should accept bad faith arguments at face value; it just means listening with an open but critical mind.

There’s no cheat code to using AI ethically, but one strategy is to examine your own motivations for using it. (Is it curiosity alone, or something else?) I talked about this in the Twitter thread that formed the genesis of this essay. To boil it down: some red flags to watch out for are profit, ego, and deception.

Are you trying to profit from other people’s uncredited, unpaid work (all the artists whose work was scraped from the web to create training datasets)?

Are you coming from a place of ego (trying to prove something or feel superior to others, perhaps those who don’t agree with you)?

Are you being deceptive (e.g. concealing that a piece of art was made using Midjourney or submitting to a magazine that doesn’t consider AI-generated content)?

If you remove the possibility of profit, then a lot of ethical cobwebs fall away. Someone who uses a free AI art generator to make character art for their own personal consumption or to share with a few friends offline probably isn’t doing much, if any, harm. But if the same person pays for a Midjourney subscription, I’d argue they’re funding harm by enabling the founders of Midjourney to profit from uncredited, uncompensated work. It’s the same question haunting everyone these days: can there be ethical consumption under capitalism? My answer: I don’t know, but you can certainly make an effort.

In a recent essay, Ted Chiang drew parallels between ChatGPT and a lossy photocopier and pointed out that some artists enjoy using such photocopiers in their process. I’m not discounting that AI could be a legitimate artistic tool in some circumstances—but we still need to tackle the issue of ownership and plagiarism.

Training an AI model on your own body of work is one way to sidestep the plagiarism issue. Award-winning SFF author Ken Liu did just that when he coauthored the short story “50 Things Every AI Working with Humans Should Know” with a neural network trained on his fiction and affectionately dubbed robo_ken. The term “coauthor” might be a stretch, since Liu disclosed only 10% of the final story’s text was generated by the AI—but clearly, in this case, AI was the tool and the human was the artist, using a computer to assist in the creative process.

Most people would find it difficult to replicate Liu’s process; creating and training a machine-learning model from scratch isn’t a common skillset today, though it might be in the future. Even if you have the technical skills, you still need a large portfolio of work to serve as a decent training set. If you have the data, then you probably also have the ability to create more art the old-fashioned way, which—for many artists—is the Whole Point.

So the pool of people with the ability and desire to train AI models on their own work is pretty small (although, thanks to Ken Liu, demonstrably not zero). But what if we could make it bigger? Imagine a service that makes it easy to build and train your own AI on your own portfolio, abstracting away the programming expertise needed for such a task. If the service is valuable, it might cost money to use, but (hopefully) its creators would only be profiting from people who explicitly consented to it and benefited in return.

Of course, the creators of this service would still run into thorny issues around copyright and people uploading art they don’t own. Turns out ethical problems don’t just go away when you ask them to. But, like asking tough questions of yourself, it’s a step in the right direction.

If there’s one thing you should take away, it’s this: artificial intelligence isn’t magic or science fiction. It isn’t sentient or actually even intelligent. It’s math—sophisticated math, but still just math. And it’s worthless without human-produced data, which means that by default, unless you’re careful, it will involve human exploitation. So by all means, be careful!

Before you try to profit from AI, understand how it works and what you’re using it to do. Tech aficionados are often guilty of trying to “disrupt” industries they don’t actually know anything about. That’s a fast track to disaster and causing harm out of carelessness. Tech moves fast, but the introduction of tech into new spaces is something that should happen slowly, and with plentiful input from people who are experts in those spaces.

I’m as guilty as anyone of getting excited and moving too fast. I made fake, AI-plagiarized book covers and posted them on Twitter. I encouraged friends to try out AI and took egotistical pleasure in explaining to them how the models worked. But what I’ve seen recently has been a sobering reminder of power dynamics and the corrupting influence of capitalism on both art and technology.

Art isn’t a zero-sum game where the winner walks away with all the cash; tech doesn’t need to be either, even if capitalists have made it seem that way. Art is awesome, and everybody deserves to make it—without causing harm to other people or their livelihoods. And everybody deserves access to the benefits of technology, too.

I love both technology and art, and what I want more than anything is to be able to feel hopeful again. Not depressed about how people are using tech to exploit and supplant art, but excited about what new things can be made possible by collusion between innovators and creatives. I want to be able to post a breathless tweet about AI art without having to delete it in shame a year later.

So, fellow tech enthusiasts, let’s do better. Let’s slow down and reexamine our motivations for creating art in the first place. The last thing we want to do is contribute to a nihilistic system that places profit before people.

I’m probably not going to make use of any AI art generators anytime soon. The ones that are out there right now don’t seem willing to publicize their datasets or make any commitments to source training data ethically. They’re more interested in protecting their profit margins and deflecting criticism with grandstanding statements that don’t acknowledge the real issues they’re creating in the art community.

But I’ll hold out hope for something better to come along, and maybe I’ll have the chance to be part of it. Maybe you will, too.

© 2023 Hana Lee