Generic selectors
Exact matches only
Search in title
Search in content
Post Type Selectors
Filter by Categories
Examined Life
Further Materials
Slush Pile
Reading Room

Dispatches from the present


What Chatbots Can’t Do


ChatGPT, the language model developed by OpenAI, has the ability to generate humanlike text on a wide range of topics. But could it write a novel? The answer is a bit more complicated than a simple yes or no.

This is what ChatGPT has to say for itself when instructed to “write the lede for an article about whether ChatGPT could write a novel.” A non-bot essayist might have opted for something with more human interest, perhaps even something exhilarating. She might, for instance, have sketched a dystopian image: a magazine office staffed by automata, a metallic hand hovering over a keyboard. But ChatGPT has no instinct for drama or scene-setting, for description or evocation. At best, its efforts could pass for an uninspired middle schooler’s term paper—adequate but unlovely.

Yet for all its plodding ineloquence, OpenAI’s innovation is the cutting-edge technology that has generated so much handwringing about the imminent obsolescence of human beings. “The College Essay Is Dead,” declares Stephen Marche in the Atlantic; “4 Jobs That Will Change (or Be Fully Replaced) by This AI-Powered Chatbot,” warns a popular tech blogger on Medium. The question of whether ChatGPT can create literature must be separated from the question of whether it could spit out ad copy or abet cheating in the classroom. My general suspicion is it cannot do any task that it’s really worth making humans do.

Is literature, then, the kind of thing that we should make humans do? Narratives that are mined primarily for their formulaic plotting can certainly be manufactured in story factories, as the likes of Marvel movies and James Patterson novels effectively already are. The former are part and parcel of film franchises that are, as Martin Scorsese put it in a despairing op-ed in the New York Times, “market-researched, audience-tested, vetted, modified, revetted and remodified until they’re ready for consumption.” The latter are not written by the figurehead but pumped out by an army of ghostwriters. Even without the aid of artificial intelligence, a sufficiently commercialized culture can automate if not art, then at least entertainment. In 1946, George Orwell was already lamenting the “machine-like … production of short stories, serials, and poems for the very cheap magazines.” Matters have only become more algorithmic in the intervening decades.

But Marvel movies and their ilk are not quite art—and art continues to be “what humans do,” as Sheila Heti puts it in her novel Motherhood. Voice, so intimately tethered to the singular sensibilities of an author, has yet to be effectively mechanized, either by machines or machinelike humans. An anti-person can only write impersonally, and ChatGPT’s attempts to compose distinctive prose have been hilariously unsuccessful. When I asked it to write a paragraph about dogs in the style of Saul Bellow, then Herman Melville, then Franz Kafka, it generated the same anonymous phrases each time. Kafka’s automated analogue could be mistaken for a teenager whining on Tumblr: “The dog, with its unassuming presence, was a constant reminder of the absurdity of existence. Its eyes seemed to reflect the confusion and disorientation that I felt, as if it too was struggling to understand the rules of this world.” These pat reflections are nothing at all like Kafka’s actual musings in “Investigations of a Dog,” which is narrated by a brooding canine who feels isolated from the rest of his species. “The others treat me with respect but do not understand my way of life,” the real Kafka has his dog lament. For its part, BellowBot writes in the same tone as KafkaBot, despite the fact that the two human writers sound nothing alike: “Dogs, they say, are man’s best friend. But what is it about these creatures that evoke such strong emotions in us?” Poor MelvilleBot is outright ungrammatical: “A dog, such an unassuming creature, yet it holds within it the power to inspire great emotion.”

Admittedly, ChatGPT will probably hone its mimicry skills over time. It works by surveying large quantities of text and determining the probability of one word following another; as it ingests ever more writing, it will regurgitate ever more fluent results. But even if it imitated Bellow or Kafka more expertly, its output would still fail to qualify as literature. The bot itself is humble enough to “know,” in some hollowed-out sense, the first reason why this is so: when asked if it can write a novel, it replies, “If you define a novel as a work of literary fiction that is original and creative, it would be difficult for me to do that because I have been trained on text that already exists and my generation is based on patterns and styles of the previous text.”

But more importantly, literature is not a mere assemblage of words that happen to appear on a page. Many philosophers of art have observed that there is something wrong with forming aesthetic judgments on the basis of testimony. Even if the most trustworthy critic in the world declares a novel beautiful, it is clear enough that I am not entitled to assert that it is beautiful until I have read it and come to the same conclusion for myself. No art lover would opt to upload correct judgments into her head and forgo the pleasure of grappling with art herself because, as the philosopher C. Thi Nguyen has argued, the purpose of aesthetic encounter is not just (or even primarily) to get things right. Instead, it is to savor the delights and demands of evaluation.

There is a parallel claim to be made, not about aesthetic appreciation but about aesthetic creation. We don’t write books (or paint paintings, sculpt sculptures and so on) just because we want to end up with them in hand, but because we want to spend our time wrangling with the phrases and wrestling with the imagery. In other words, we want to suffer through the challenges of authorship, and unless we are James Patterson, it would defeat the point to delegate someone else to do it for us.

And similarly, we read not to end up with information lodged in our brains, but in order to come face to face with an author’s skirmishes with language. Indeed, we read novels, rather than textbooks or user’s manuals, because we are not in the business of extracting propositions but in the business of effecting intimacy with another live intelligence. Literature is not (only) a conveyer of information but a locus of communication, and we cannot communicate with an inanimate mechanism, whirring its insensate way through text it does not even comprehend. For this reason, we could only ever really care about words that have been deliberately placed on a page by another person. Books, the German Romantic novelist Jean Paul once wrote, are “thick letters to friends.” Who would want to correspond with the void?

Photo credit: Mirko Tobias Schäfer (Flickr, CC / BY 2.0)