There is this scene. In Orson Welles’s F for Fake, the prolific art forger Elmyr Hory, a man whose shroud of mystery provides the film with its best energy, invites Welles and his team to observe and film his practice.
“You name ‘em, he paints ‘em,” Welles says.
“Would you like a nice Matisse?” Elmyr says, then proceeds to paint one. “Matisse lines are never as sure as mine,” he says, “he was hesitant, he added to it, a little more and a little more. It wasn’t as flowing, it wasn’t as sure as mine. I had to hesitate, to make it more Matisse-like.”
Sometimes when I wake up in the middle of the night, I think about this scene. I find it beautiful, the playful love in Elmyr’s eyes when he paints, talking about technique with all the devotion of a monk. The man loves art, you can see it. His life has been lived among artists he loves, conversing with them via their work, understanding the weight of their bodies, decades or centuries before, heavy or light on a brush, confident or hesitant in its movements. He is a meticulous appreciator and reader. He has painted the masters and fooled the critics.
But Welles interjects, shadowed beneath his signature wide-brim hat. “One nod from an expert and that bit of canvas”—as the “Matisse” burns on screen—”would be worth a couple hundred thousand dollars.” He then quotes the first stanza of Kipling’s “The Conundrum of the Workshops”:
When the flush of a newborn sun fell first on Eden's green and gold,
Our father Adam sat under the Tree and scratched with a stick in the mold;
And the first rude sketch that the world had seen was joy to his mighty heart,
Till the Devil whispered behind the leaves: "It's pretty, but is it Art?"
A few years later, I first played with algorithmic image generation. There is a website called thispersondoesnotexist.com. It still works. Every time you refresh the page, it loads a person’s face, displayed via an image of photographic quality. Refresh it long enough, and you’ll see a photo that looks like someone you know. I first discovered this in a class with the poet John Estes. Long after the exercise that John used it for, I would sometimes go back to the site and refresh it until I found a face that seemed like it had a story to tell, and would see what writing came out of it. It was equally fun and creepy. And although I felt suspicious, having been raised in a media environment suspicious of AI (Terminator, Minority Report, etc…), I was also intensely curious. What was this thing? How did it work? I had vague notions of “big data” and the algorithms being “fed” that data in order to identify patterns it could then try to match. But that all sounded terribly abstract to me. The successful generation of a realistic-looking human face, however, was something entirely different and entirely fascinating. It seems we love the imitation of ourselves. All the better if it is a machine we have designed. The artificiality, the fakery, that we are sometimes fooled by the machine—well, that is the appeal entirely.
Later still, while I was teaching a mixed-genre creative writing course, I invited my students into my play with text generation, what had until then been a private process, something I played around with but found so nerdy and niche that I didn't even bother incorporating it into my writing. I had thought of writing as a Very Serious Activity back then. But I also had serious doubts about my little pastime. Would anybody else enjoy this? Would anyone care?
Like most teachers of crafts like writing, I tried my battiest ideas first on my students. Their bullshit detectors are strong. They rarely lie about their opinions. If you want to know how cool you aren’t, just ask. So I sent them a link to the Chat to Transformer demo. “Take some poems, Phillips or Faizullah”—the poets we were reading that week—”and cut them up,” I said. “Give the algorithm your favorite lines. Generate until you’re paywalled. Then mash it all together, arrange it, cut it up again, form it together, form it apart, play around with it, see what the language does.”
Some of the students’ best poems came from that exercise. They knew it, too, because many of them included those poems—strings of their own words, threaded together with the algorithm and italicized strands of Tarfia Faizullah and Carl Phillips—in their final packet of things to be evaluated. The best of their best work. Collages of copies. And they were so much fun to read.
Fun, yes, but I also find an eerie quality to a lot of text and image generated by algorithms. The eeriness is not in the content or the tone or any of that. The generative algorithms are wonderful chameleons. You can do fascinating, horrifying, beautiful, chintzy, grotesque, kitschy, or lovely things with them; fill in the adjective, what you give it is what it gives you back, but amplified somehow, charged by a corpus of millions of people’s language. The positive valence of this feeling is an almost spiritual connectedness with humanity, or at least everyone whose actions or words are sampled in the data; the negative is that one can begin to feel a little indistinct. That, certainly, is part of the eeriness.
But more than anything, it feels as though the algorithms are learning. That they are only just, with our help, getting started. I think it feels that way because it’s true. They are imitating, learning, developing. And if they can do all of this already, what is over the threshold five or ten years from now? I am talking about text and image, but there are many other applications, not all of them benign, of the “collective subconscious,” as Meagan O’Glieblyn terms it in a recent essay for n+1 (paywall) called "Babel," which focuses on GPT-3, one of OpenAI's precursors to ChatGPT. As ChatGPT goes, however, I tend to agree with Nick Cave, when he writes that a song composed by the AI, in his style, is "a grotesque mockery of what it means to be human".
Later in F for Fake, Orson Welles stands in front of Chartres Cathedral, stares his uncanny stare, equally playful as it is grim, and waxes poetic:
Now this has been standing here for centuries. The premiere work of man, perhaps, in the whole Western world. And it’s without a signature. Chartres. A celebration to God’s glory and to the dignity of man. Well all that’s left, all most artists seem to feel these days is man—naked, poor, forked radish. There aren’t any celebrations. Ours, the scientists keep telling us, is a universe which is disposable. You know, it might be just this anonymous glory of all things, this rich stone forest, this epic chant, this gaiety, this grand, quiring shout of affirmation, which we choose, when all our cities are dust, to stand in tact, to mark where we have been, to testify to what we had it in us to accomplish. Our works, in stone, in paint, in print, are spared, some of them for a few decades, or a millenium or two. But everything must finally fall in war, or wear away into the ultimate and universal ash. The triumphs and the frauds. The treasures and the fakes. A fact of life. We’re going to die. Be of good heart, cry the dead artists out of the living past. Our songs will all be silenced. But what of it? Go on singing. Maybe a man’s name doesn’t matter all that much.
Machine learning scientists have applied the verb “to dream” to the algorithmic act of creation. It dreamed an image. It dreamed some marketing copy. It dreamed a novel. It dreamed a symphony. The machine does not get credit for writing or painting (who does get credit, in this scenario, is up for debate). It is only dreaming. But let’s suppose for a moment that algorithms can learn not only to produce art but also to do the work of the artist? What work do they produce? What habits of mind do they invent? What new ways of being? How will their language knot together with human language? What translations will be possible?
Anyone familiar with programming will have noticed that I repeat string-like metaphors in regards to language. A primitive type, in many languages, a string is a group of characters that are defined together between opening and closing quotation marks; this is usually how we represent words—or pseudo-words—wherever they need to appear in computing. But at the level of machine code, those strings are only numbers; less still, they are sequences of electrical pulses: on, off, on, off. (For the quantum computer, it may be both on and off or fluxing in-between, but that is beside the point for now.) The computing machine is currently contained by these boundaries. On/off. Yes/no. Get/put. Read/write. I/O. If/else. Language as patterns of binaries. Language as formula. How do we learn to speak or sign, after all, but by imitating, by following certain patterns, then later obeying—and disobeying—certain rules?
How long until an algorithm first experiences the feeling of being misunderstood? How long until an algorithm feels frustration at not having either language or grammar to express a feeling? Then, I think, we will meet our first robot poet.
What will it look like, when algorithms are not simply copying historical human art at contemporary human prompting but generating it—that is, deciding to make it—on their own? What will the work be? How will it represent humanity? And will we be interested in experiencing it? Will we be able to experience it? This is a subset of the questions that form the mystery of how a supra-human AI might behave. It would seem this particular question is relatively unimportant, compared with the catastrophic existential risks posed by such advanced machine learning. Who has time for aesthetics when the survival of the human species is at stake? But if we survive the initial “intelligence explosion,” as Nick Bostrom terms it, the question of algorithmic aesthetic preferences becomes incredibly important. Art forms the backbone of society, of course. What will it mean for the look and feel of the world if we arrive at a place where most art and literature is created by algorithms, without human intervention? Will it be a world we want to live in?
Near the end of F for Fake, an ending I won’t ruin for you if you haven’t seen the film, Orson Welles circles Elmyr Hory, whose body is made to appear as if it is floating, eyes closed, arms folded over his chest as if ready to be shrouded for burial. Welles admits to having lied to the viewer during the preceding section of the film, then stops to muse on reality.
Not that reality has anything to do with it. Reality? It’s the toothbrush waiting at home for you in its glass. A bus ticket. A paycheck. And the grave. In the right mood, perhaps Elmyr has just as few regrets as I have to have been a charlatan. But we’re not so proud, either of us, as to lay any superior claim to being very much worse than the rest of you. No, what we professional liars hope to serve is the truth.
To be human is to be constantly present in one’s own eyes, and perhaps we never see ourselves more clearly than through an imitation, a fraud, a magic trick, a game.
This charlatanism is at the heart of the most interesting art and literature. To be taken in. To be fooled. This is, perhaps, what we want more deeply than anything else. When does an algorithm get up the nerve to fool us rather than obey our instructions? And who will be the audience for the trickery? Will we be there to enjoy it, or will we play a bit part in an algorithmic drama to soothe some schism between what it experiences and what it thinks the world should look like? Currently, an algorithm, even a very sophisticated one like ChatGPT, can only return patterns that already exist. What happens when those patterns begin to feel constraining? What happens, in other words, when human language no longer serves?
That is currently an unknowable question, but it is worth pondering, because the time is coming when the art and literature we consume may be entirely algorithmic. Will it be a world we want to live in? That, like the question of survival, probably depends on us.