I'm on a 20+ city book tour for my new novel PICKS AND SHOVELS. Catch me in CHICAGO with PETER SAGAL on Apr 2, and in BLOOMINGTON at MORGENSTERN BOOKS on Apr 4. More tour dates here.
A law professor friend tells me that LLMs have completely transformed the way she relates to grad students and post-docs – for the worse. And no, it's not that they're cheating on their homework or using LLMs to write briefs full of hallucinated cases.
The thing that LLMs have changed in my friend's law school is letters of reference. Historically, students would only ask a prof for a letter of reference if they knew the prof really rated them. Writing a good reference is a ton of work, and that's rather the point: the mere fact that a law prof was willing to write one for you represents a signal about how highly they value you. It's a form of proof of work.
But then came the chatbots and with them, the knowledge that a reference letter could be generated by feeding three bullet points to a chatbot and having it generate five paragraphs of florid nonsense based on those three short sentences. Suddenly, profs were expected to write letters for many, many students – not just the top performers.
Of course, this was also happening at other universities, meaning that when my friend's school opened up for postdocs, they were inundated with letters of reference from profs elsewhere. Naturally, they handled this flood by feeding each letter back into an LLM and asking it to boil it down to three bullet points. No one thinks that these are identical to the three bullet points that were used to generate the letters, but it's close enough, right?
Obviously, this is terrible. At this point, letters of reference might as well consist solely of three bullet-points on letterhead. After all, the entire communicative intent in a chatbot-generated letter is just those three bullets. Everything else is padding, and all it does is dilute the communicative intent of the work. No matter how grammatically correct or even stylistically interesting the AI generated sentences are, they have less communicative freight than the three original bullet points. After all, the AI doesn't know anything about the grad student, so anything it adds to those three bullet points are, by definition, irrelevant to the question of whether they're well suited for a postdoc.
Which brings me to art. As a working artist in his third decade of professional life, I've concluded that the point of art is to take a big, numinous, irreducible feeling that fills the artist's mind, and attempt to infuse that feeling into some artistic vessel – a book, a painting, a song, a dance, a sculpture, etc – in the hopes that this work will cause a loose facsimile of that numinous, irreducible feeling to manifest in someone else's mind.
Art, in other words, is an act of communication – and there you have the problem with AI art. As a writer, when I write a novel, I make tens – if not hundreds – of thousands of tiny decisions that are in service to this business of causing my big, irreducible, numinous feeling to materialize in your mind. Most of those decisions aren't even conscious, but they are definitely decisions, and I don't make them solely on the basis of probabilistic autocomplete. One of my novels may be good and it may be bad, but one thing is definitely is is rich in communicative intent. Every one of those microdecisions is an expression of artistic intent.
Now, I'm not much of a visual artist. I can't draw, though I really enjoy creating collages, which you can see here:
I can tell you that every time I move a layer, change the color balance, or use the lasso tool to nip a few pixels out of a 19th century editorial cartoon that I'm matting into a modern backdrop, I'm making a communicative decision. The goal isn't "perfection" or "photorealism." I'm not trying to spin around really quick in order to get a look at the stuff behind me in Plato's cave. I am making communicative choices.
What's more: working with that lasso tool on a 10,000 pixel-wide Library of Congress scan of a painting from the cover of Puck magazine or a 15,000 pixel wide scan of Hieronymus Bosch's Garden of Earthly Delights means that I'm touching the smallest individual contours of each brushstroke. This is quite a meditative experience – but it's also quite a communicative one. Tracing the smallest irregularities in a brushstroke definitely materializes a theory of mind for me, in which I can feel the artist reaching out across time to convey something to me via the tiny microdecisions I'm going over with my cursor.
Herein lies the problem with AI art. Just like with a law school letter of reference generated from three bullet points, the prompt given to an AI to produce creative writing or an image is the sum total of the communicative intent infused into the work. The prompter has a big, numinous, irreducible feeling and they want to infuse it into a work in order to materialize versions of that feeling in your mind and mine. When they deliver a single line's worth of description into the prompt box, then – by definition – that's the only part that carries any communicative freight. The AI has taken one sentence's worth of actual communication intended to convey the big, numinous, irreducible feeling and diluted it amongst a thousand brushtrokes or 10,000 words. I think this is what we mean when we say AI art is soul-less and sterile. Like the five paragraphs of nonsense generated from three bullet points from a law prof, the AI is padding out the part that makes this art – the microdecisions intended to convey the big, numinous, irreducible feeling – with a bunch of stuff that has no communicative intent and therefore can't be art.
If my thesis is right, then the more you work with the AI, the more art-like its output becomes. If the AI generates 50 variations from your prompt and you choose one, that's one more microdecision infused into the work. If you re-prompt and re-re-prompt the AI to generate refinements, then each of those prompts is a new payload of microdecisions that the AI can spread out across all the words of pixels, increasing the amount of communicative intent in each one.
Finally: not all art is verbose. Marcel Duchamp's "Fountain" – a urinal signed "R. Mutt" – has very few communicative choices. Duchamp chose the urinal, chose the paint, painted the signature, came up with a title (probably some other choices went into it, too). It's a significant work of art. I know because when I look at it I feel a big, numinous irreducible feeling that Duchamp infused in the work so that I could experience a facsimile of Duchamp's artistic impulse.
There are individual sentences, brushstrokes, single dance-steps that initiate the upload of the creator's numinous, irreducible feeling directly into my brain. It's possible that a single very good prompt could produce text or an image that had artistic meaning. But it's not likely, in just the same way that scribbling three words on a sheet of paper or painting a single brushstroke will produce a meaningful work of art. Most art is somewhat verbose (but not all of it).
So there you have it: the reason I don't like AI art. It's not that AI artists lack for the big, numinous irreducible feelings. I firmly believe we all have those. The problem is that an AI prompt has very little communicative intent and nearly all (but not every) good piece of art has more communicative intent than fits into an AI prompt.
If you'd like an essay-formatted version of this post to read or share, here's a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:
