The Literature of the Improbable

 In an episode toward the beginning of Donald Barthelme's 1967 postmodern novel, Snow White, the titular fairy tale protagonist confronts the seven men with whom she shares a ménage, saying: "Oh I wish there were some words in the world that were not the words I always hear!" (Well, I guess this isn't a confrontation precisely, so much as a request phrased as complaint; something on the order of "will no one rid me of this turbulent priest?") The seven men soon fall over themselves trying to supply words and phrases that fall out of the ordinary, and thus could not be the ones she is used to hearing: "fish slime," says one; "Injunctions," says another. "Left sucking the mop again" is a third phrase that gets thrown around, and this one sticks—it comes up at least twice more in the novel, where it appears to mean something like "left behind the eight ball." 

Now, I bring this up because I am persuaded that we are all about to experience some version of Snow White's complaint, in the next few years. We are living on the eve of a millennial deluge of over-familiar words. The reason is: AI language models. Humans have now built machines that are astoundingly gifted at guessing statistically-likely correlations of words, and can therefore generate the most plausible of texts. So gifted are they at this ability, that one can start to recognize the product of a generative AI language model by the sheer probability of its words. A human reading it gets the vague sense, "this feels like a genuine original text; and yet it is familiar somehow; it reads like precisely what you would expect to get, if you fed all the internet into a machine and asked it to reproduce the average content of the whole." And since the AI can produce this content in seconds, there is bound to be a lot of it soon. 

That is why I say we will suffer from Snow White's condition. The words around us will shortly be the ones we always hear, and we will find ourselves wishing for that which cannot be so easily guessed; for that which does not statistically correlate to itself. But here, precisely in this impasse, is perhaps the hope for our artistic redemption as a species. 

The rise of AI, after all, is not the first time that the creative professions have confronted the threat of displacement by non-human technology. Now is not the first time, that is to say, we have been forced to consider the position of "the work of art in the age of mechanical reproduction," to borrow Walter Benjamin's phrase. When the photograph and the daguerreotype emerged in the nineteenth century, visual artists had to confront the fact that oil painting was no longer the only or the best way to approximate a realistic visual representation of the world. If there was to be any justification left for their craft, therefore, it would have to be found in something other than a unique capacity to mirror nature. Thus, the rise of anti-realistic schools of painting followed: impressionism, expressionism, cubism, abstraction, and all the rest. 

Visual artists managed to make a virtue of necessity, in short. Instead of merely bemoaning the tide of technological change—instead of breaking through the doors and windows of the photograph shops and setting alight their equipment—they found a way to push out into other realms that technology had not conquered, and could not so easily reproduce. 

Plainly, creative prose writers will need to find a means to do something similar. But what could that be? What is it that human writers can produce that the new AI language models would not so easily mimic? 

The answer, I reply, must lie in the literature of the improbable. Think about it. What is it that distinguishes AI prose? It is the machine's incredible knack for guessing highly-correlated combinations of words, based on mapping the statistical relations of all known words on the internet. Thus, as with the portrait painters put out of work by the rise of the photograph, the economy will soon no longer have need of human writers to generate plausible-sounding text. The ability to take a piece of prose—advertising or website copy, say—and work that magic upon it that makes it seem "polished," or "professional," or simply "like everything else," as if it had existed always, and no one had written it at all—will no longer be as highly prized as it was when I was coming up through the ranks of the digital communications profession. Instead, something else will be called for: the ability to generate the improbable. 

It is not just my imagination, by the way, that here—in the realm of the statistically-unlikely combination—is a sort of prose that remains the domain of the human creator. My sister was telling me the other day that the early software that has been developed to catch the use of chatbots in student writing operates by scoring the prose on how statistically-guessable it is. That which is comprised of a highly-correlated set of words is pegged as probably written by a machine. "But this seems unfair," we both agreed, when contemplating it. Because the software could also just be punishing those students who have learned how to write in the most probable of ways—who have mastered the art of "sounding like everyone else," in short. And is this not precisely the skill that has been demanded of generations of copy writers and communications professionals up to this point—to stifle any flickerings of a florid prose style; to leave only the most familiar and inconspicuous of phrases in their place? 

But if this long-cultivated skill may soon become tragically otiose, its death also points the way to how creative prose may be revived. Its future may lie precisely in the generation of those combinations of phrases the machine would never guess: in the unfamiliar and wholly outside-the-ordinary concatenation of words. "Fish slime," "injunctions," and "sucking the mop," humble as they seem, may point the way to our one and only hope of artistic salvation. What we will need to master is the literature of the improbable—the art of the exception. 

I made a similar point about science in an earlier post, in which I suggested that the future of scientific invention in the age of AI might lie in "'pataphysics"—Alfred Jarry's humorous discipline that distinguished itself from the conventional sciences because, whereas those other logoi devoted themselves to the study of the rules and laws of nature, 'pataphysics was interested precisely in the lawless "exceptions"—the anomalies, in short. 

Something similar could be said of the most promising future direction for literature: if the chatbots have become gifted at predicting the statistical laws of prose; then the human-generated prose of the future, in order to open new fields for cultivation, will have to concern itself with the statistical exceptions of prose: the unlikely word or turn of phase; the unpredictable combination of events and scenes that defy convention. 

Of course, this is not an entirely new idea. We have had a literature of improbable combinations already from Raymond Roussel to Harry Mathews. The avant garde has made a point of generating statistically-unlikely combinations of scenarios, events, and characters for more than a century, and it is not my intention here to merely propound one more manifesto of surrealism. Nor of literary postmodernism, for that matter. 

Likewise, the crusade against over-used words and phrases is also hardly a novel approach in the ranks of high-minded literature. Writers who aspire to immortal fame have long tried to rid their prose of the predictable combination of words and the over-familiar phrase, metaphor, or concept. Perhaps, then, all we are calling for here is just another version of Martin Amis's (may he RIP) "war against cliché," or of Flaubert's campaign against "received ideas." 

Even if my proposal is not entirely novel, though, it is fair to say that the stakes are higher and the urgency is greater for it than ever before. The statistically-guessable prose is no longer to be ruled passé merely because anyone can do it, given enough time and practice and mastery; it is to be rendered outright otiose by the fact that any machine will soon be able to do it, and in a micro-fraction of the time, too. If prose writers want to find something that they can do that the machines are unlikely to reproduce—just as visual artists had to go searching for a style of painting that a photograph could not replicate—then this may be the sole avenue down which to stroll. 

Donald Barthelme's work may itself be a place to start, in seeking for exemplars, and not only in the brief section of his novel where the dwarves are shouting odd phrases. 

To be sure, one of Barthelme's gifts is as a mimic—his stories are a jazzy bricolage of the various prose styles that flutter ambiently around us, in modern society, and Snow White is no exception: here, one finds flawless renditions of the patriotic harangue, the courtroom cross-examination, the therapeutic self-help bromide, the Madison Avenue advertising copy, and countless more. This capacity for parody is itself, of course, a kind of AI skill: it comes from having an ear for statistical correlations. But the combination of these parodies, the melding of high and low diction, the fact that this is all taking place in a modern-day fairy tale set in a city with seven amorous dwarves, a villain named Hogo de Bergerac, and a sitting U.S. President who periodically concerns himself personally with the events the novel describes, surely make this body of prose sufficiently unlikely to place it into the realm of the improbable. 

And so, apple-monstrance; verdigris; archaeopteryx; lemon custard—oh, who am I kidding, there's probably some subconscious process of association at work here that the machines will one day be able to guess!