Art-ificial Intelligence
Generative AI, writing, art, choices, and Ted Chiang
Please hit the ❤️ “Like” button at the top or bottom of this article if you enjoy it. It helps others find this article.
Ted Chiang is, in my estimation, the best active writer in science fiction.
He wrote the story that was turned into the hit sci-fi movie Arrival. He writes short-form fiction, and his two collections (Stories of Your Life and Others and Exhalation) are both jaw-droppingly amazing.
His prose is delightful, his stories thought-provoking, and his introspection deep. There are no weak stories in his portfolio. When I read his writing, I just can't believe how good he is at it.
I have no love for AI hype. I'm a data scientist. I work with machine learning and "AI" at my day-job. I've managed teams implementing software features using the latest generative AI models. I can appreciate generative AI, but there's a reason I don't write an AI blog. It would be like an accountant having a blog about spreadsheets. I see enough of it at work, it's not something I want to spend my free time on.
When I hear the hype, it's painful. The AI space is full of snake oil and bullshit. Exaggerations are the norm. People are trying to cash in on the excitement of people that don't have a firm grasp of the technology. Speculations run rampant with no bearing on reality.
So it pains me so much to say that, in a recent op-ed where Ted Chiang takes on the AI hype regarding art, he misses the mark.
Art and choices
Chiang sees art as about making a lot of choices—big choices (what general topic to write/draw) and little ones (specific word choices or brush strokes). Crucially, he sees artistry as the interrelation between those large-scale choices and the small ones.
Chiang frames it this way because he is taking aim at a particular view: the view (pushed by the marketing departments of big tech companies) that, with AI, we can "unlock our creativity" by turning our ideas into art. Simply write a short prompt and BOOM, you've made your dream novel.
According to this AI-hype view, we all have latent creativity waiting to be tapped. We've all had great ideas for a story or illustration, but we haven't been able to create that art because we haven't had the time to hone the skill. Hence, AI is unlocking that potential by removing an obstacle to creating art.
Chiang argues that this is a naïve view of art. Writers constantly complain about being approached by people who "have a great idea for a book". The problem is, having a single (high-level) idea isn't the same as art. Ideas are cheap and easy. Where the rubber hits the road is in fleshing out that idea. The small-scale choices breathe life into the large-scale ones.
If this was the limit of Chiang's critique, I wouldn't be writing this. For criticizing some AI marketing hype, this is all fine. But Chiang wants to go much further, to claim that because art is about choices, AI fundamentally can't add to it. And he sees this applying not just to high art, but writing meant to entertain as well.
Comforting mental models
I write fiction. It would be really comforting for me to have some golden rule for why the particular way I write today is the way it must be done and that there will always be a commercial market for it.
This is the allure of the mental models that critics of AI generation employ.
Central to Chiang's view is that AI can't make choices. He suggests all AI can do is take the average of choices other writers have made, or mimic the choices of one particular writer. He claims the result is bland or derivative, and implies it couldn't be otherwise.
I agree that current generated content generally sucks. But Chiang is misrepresenting what you can do with AI. You can prompt them to generate text by mashing up the styles of two authors, or input a sample of your own writing and ask it to match your style, or ask it to write like an author but punch up certain aspects.
These ideas map pretty well onto cognitive science models of creativity. Combinational creativity is the combining of two things previously not combined. Transformational creativity is taking something and transforming some aspect of it to generate something new. It's easy to instruct generative AI to do these types of things, so it isn't obvious what comes out must always be derivative.
There isn't some simple reason generative AI fundamentally can't be used to create something creative, entertaining, or that could meaningfully be called art. To use Chiang's framing, there's no reason AI can't contribute to or make some of the choices involved in art.
We can just look around the world at what's going on. People are already using AI to generate large parts of creative writing that is clearly entertaining since many people are continuing to be fans of the authors putting it out. Rie Kudan used AI generated text in her novel that won one of the top Japanese literary prizes. A recent survey by the Society of Authors found 20% of fiction and 25% of non-fiction authors have used generative AI for their writing.
This is happening even though the current quality of AI generated writing sucks. We all know it sucks. Yet some writers are still finding it useful.
There's no simple rule that says AI can't be used to make art. Chiang wants to paint a picture where art takes choices and humans must make those choices, so AI can't meaningfully contribute to art. But there just doesn't seem to be a good reason to think that.
Instead, I think there's a moral judgment going on here.
Chiang, along with many other critics of AI-generated content, see these models as "laundering" the work of others. Since the training data for generative AI comes from human writers, he sees it as plagiarism.
The plagiarism model of what generative AI is doing is one view. The opposing view is that these models learn from writing just like humans do. If it’s fair for a human to learn how to write from reading others, AI should be able to as well.
I can see the appeal of both views. But both are relying on analogies that aren't quite right. The type and scale of learning generative AIs do certainly isn't just like how human learning works. But to claim it’s plagiarism seems weird—plagiarism means copying from a specific work. That AI is plagiarizing, but it isn't copying or paraphrasing from a single source, differs greatly from our usual conception of plagiarism.
We reach for analogies because we want to reduce the ambiguity of what this new technology is doing to something we understand and have a good mental model of. We want to know the limits of what it can do and come up with the right moral framing for it. But we should disregard the impulse to reach for the simple answers.
Maybe the place of AI in the future of writing is as a tool writers use. It has different strengths and weaknesses that can help augment human writers. I don't have any special insight here. My point is just that we should exercise some humility before claiming surprising hard limits on what generative AI can do based on weak mental models or analogies.
But there's a much more interesting analogy that Chiang draws in his critique.
Photography and the future of AI art
Chiang mentions photography as an example of a past technology that seemed to threaten creativity. On the surface, it seems like it removes the need for human visual artists. We can just snap a photo. But instead, it created a new art: photography. It might not have been immediately obvious, but photography involves a myriad of choices that a well-trained photographer can make to create art. Chiang points out photography requires many big and little choices, but can't imagine how AI could (in his defense, he's mostly talking about it in the context of creating an image or story).
But the AI-hype vision of a short prompt generating a complete novel is just one picture of how one could use AI to make art. And it's a boring, uninspiring one. It would be like seeing a camera and thinking you could only use it in one static position to take portraits of people sitting in one position.
What's interesting about generative AI isn't its ability to generate static text or images. It's the interactivity. That's the component that opens up possibilities for alternative choices that can create art.
As an analogy, let's look at one of the common ways companies use generative AI: Retrieval Augmented Generation (RAG). Companies use these systems in chatbots (it's also, I expect, how Google's Gemini serves up answers to questions at the top of your search query).
The overall idea is pretty simple: let's say you're a user with a question about how to use a product. Documentation exists, but it's large and hard to navigate. So you ask a question to the RAG (displayed to you as a chatbot). Behind the scenes, a retrieval system, based on the normal search algorithms we've all been interacting with for years, grabs the top hits from the documentation library. Those documents, along with your question, are fed to a generative AI. The generative AI then uses that information to provide you with an answer.
RAGs work pretty well. The retrieval system providing relevant documentation overcomes the tendency for generative AI to "hallucinate" (make up answers), and the generative AI can often give much more concise and accessible answers than you would get trying to digest the full documentation returned by a search engine.
The generative content returned by a RAG is based on three things: An initial prompt the engineer setting up the RAG gives the generative AI (like "You are an expert on this system, give polite and clear answers, do not tell the user you're planning to exterminate humans", etc), the user's question, and the documentation. The engineer setting up the system puts in a bunch of additional constraints (like logic on what to do if no relevant documentation exists, or to prevent various other forms of inappropriate responses).
Companies use RAGs for a pretty boring purpose: as slightly less frustrating chatbots that hopefully can actually answer your questions about a product or service. But they're interesting for another reason: they're a demonstration of what generative AIs allow you to do beyond the boring creation of static text. They can create interactive experiences where new text is generated based on a mix of what the creator puts in, the user puts in, and any other accessible information.
It's easy to think of entertaining and possibly artful interactive experiences you could create with this.
Choose-your-own-adventure novels provide a different experience than normal novels because they allow some small amount of interaction between the reader and the book. Imagine instead a choose-your-own-adventure where the creator writes a broad outline of the world the story takes place in, the rules of it, the types of endings, and the introduction.
But they don't have to anticipate every possible thing the user can do. They can give guidance, examples of scenes to use, characters that might appear. They could write a whole book-worth of information fleshing out the world and events going on in it. Then the generative AI can riff on this, based on what the user inputs.
Something like this already exists with AI Dungeon, an app that uses generative AI to create interactive stories. I'm not claiming it's art. But given how new the technology is and the current limitations of it, it's an interesting glimpse into potential artistic uses of generative AI as this technology improves.
This has more obvious applications in mediums that are already highly interactive, like video games. Generative AI can provide richer dialogue for non-player-characters in the game to bring life to a medium that often struggles with feeling too scripted. Video games are already a highly interactive art form (yes, video games can be art), and generative AI can become a new component of them.
I'm not claiming that art using generative AI will look exactly like what I just described. This is just a rough sketch of one direction. But just like with photography, we shouldn't assume that the artistic choices opened by a new technology need to conform exactly to the art forms of the past. We're in the early ages of high-quality generative AI. Lots of creative people are actively using them. Let's not assume nothing interesting can come of that.
Existential crises for writers
Where does this leave us writers?
I think the big concern for writers is that, by making creative writing easier to generate, generative AI will further crowd the already small commercial niche of fiction writing. Writing is already cheap because so many people want to be writers, and this would exacerbate the problem.
AI-generated content is already doing damage to the markets for fiction writers—some magazines have reported being flooded with low-quality generated stories, putting a strain on their already limited resources to sift through the stories being submitted.
I don't have any special insight into how this will play out. For now, stories wholesale generated by AI suck. Even if the tools are useful to writers at this point, they're not game-changers by any stretch of the imagination.
People like reading things from their favorite writers. There's a personal connection part that's important.
Maybe those factors won't change, and the literary world will continue on much as it is now—most writers doing it practically for free while a privileged few make a living doing it (and an extremely small number get rich).
Or maybe in 20 years the art world will look different. Maybe there won't be a commercial landscape for stories like there is now, and the already small "privileged few" will be reduced to even fewer.
Writing—and art more generally—create what economists call "positive externalities". We are all enriched by living in a culture with artists, but not all of the value they create returns to them. This is why things like art and culture grants make sense—we should collectively pay for the things that enrich us all. In an age where the commercial use of art becomes even more squeezed, it might become even more important to have these sorts of programs.
But even if we don't collectively recognize the need to fund art, art will still exist. As Chiang points out in his essay, there are reasons beyond the end-product for creating art. It's communication. It's a way to explore and express thoughts and emotions. If you subtract out the commercial side of things, there's still a need for art created by and for humans.
Instead of declaring AI can't do specific things, we should be realists. Maybe art will evolve or maybe it will stay more or less the same. Regardless, we can choose to embrace generative AI and explore the new technology with the open, creative minds of artists. We can continue to create in whatever form for the expression and communication it provides rather than the commercial viability, and take joy in honing our craft as storytellers and artists. And we can work as a society to pay back those artists that create positive externalities, regardless of what tools they use.
Please hit the ❤️ “Like” button below if you enjoyed this post, it helps others find this article.
If you’re a Substack writer and have been enjoying Cognitive Wonderland, consider adding it to your recommendations. I really appreciate the support.
This is really well articulated, and I’m particularly interested in the interactivity component of AI in writing. Maybe this is unpopular to admit, but I often use AI as a conversational companion (I think best when in dialogue with others) to help suss out what I’m trying to say. I hardly ever use the actual words AI produces, but it’s more that the back and forth will spark something that leads to further clarity. When writing about my own past and history, it has even felt… dare I say… kind, like there was someone there listening to me and echoing me as I did the difficult work of excavating the past.
A lot of your commenters are using an analogy with photography to say that AI fiction will not be all bad. But extend the analogy and think about what happened to art in the first hundred years of photography.
Mid-19th century art was mostly representational and becoming more life-like until the Impressionists zigged left and conveyed a mood as much as capturing an image. Images were for cameras. As the 20th century went on — Seurat, Duchamp, Kandinsky, Mondrian, Pollock — art got further and further from what the artists of the mid-19th century would have called art.
While photography took most of the business of representational art, painters had to create new genres to even have a place where they could contribute. There’s a good side to this: photography is beautiful and gave us new ways to speak with an image and we got Picasso and Chagall. But there were also 100 years where more traditional artists lost their calling while their audience flocked to the new technology which was cheaper and more available. It’s nice that we have photography, but it took a long time to get to the point where photography gave us back what we’d lost.