Did Donald Barthelme Anticipate AI?
Barthelme is a quintessential writer of the twentieth century, looking Janus-faced to both the past and future, and with a third eye turned inward. — David Gates
Given a word, what should the next word be?
Donald Barthelme is a writer’s writer. Francine Prose, Lydia Davis, Dave Eggers, and George Saunders have gushed about him. Barthelme once invited — get ready for this list — Thomas Pynchon, John Barth, William Gaddis, Robert Coover, John Hawkes, William Gass, Kurt Vonnegut, Walter Abish, and Susan Sontag to a dinner and everyone showed up except Pynchon. David Foster Wallace said Barthelme’s story “The Balloon” made him want to be a writer. (Wallace 2012)
Barthelme’s writing is strange but also funny. He was intensely curious about the world around him and plumbed its depths to find new ways to write:
Barthelme’s project of restoring freshness to language…led him to ragpick words, phrases, tones of voice, and modes of diction from the obscure and neglected past…from the surreal specialized lexicons of technology, philosophy, and even the military. (Gates 2003)
The New Yorker published “The Explanation” on August 27, 1968. Barthelme was 37. It begins like this:
Q: Do you believe that this machine could be helpful in changing the government? (Barthelme 1968)
In 1968, machines were on peoples’ minds — especially thinking machines. Was it on Barthelme’s mind? I don’t know. But I think so.
A Turing/Hilbert interlude
Alan Turing died in 1954, fourteen years before “The Explanation” was published. Barthelme was 23. The idea of thinking machines predates Turing though. Leibniz, co-inventor of calculus, built a calculating machine in 1673. He wondered if machines could determine truth values — that is, determine if a statement is true or false. How would that language look? Probably like an algorithm.
Algorithms. In 1900, in a monumental address to the International Congress of Mathematicians, David Hilbert presented 23 unsolved math problems that could be guideposts for mathematical thought in the coming century.* Here’s problem 10:
Given a Diophantine equation with any number of unknown quantities and with rational integral numerical coefficients: To devise a process according to which it can be determined in a finite number of operations whether the equation is solvable in rational integers.
“Process” has been taken to mean “algorithm.” Alan Turing heard about all this — and about problem 10 — in Max Newman’s class at Cambridge. One day Newman ended his class with a question that would change Turing’s life:
At one point [Newman] posed his class a question: could the provability of mathematical statements be discovered by a mechanical process? (Owen 2012)
The phrase “by a mechanical process” has been credited as the reason Turing became obsessed with machines. (Hodges 2014)
Barthelme was deeply curious but also deeply anxious about being influenced — too much — by other writers. Could he have seen the machine world as a wellspring of ideas and syntax? Not Gertrude Stein’s automatic writing, but something more and wholly unique — syntax not of the subconscious mind but of the unconscious machine mind. A mechanical process.
We know he was aware of Gertrude Stein. She is on page one of his book Not-Knowing:
Joyce, Gertrude Stein, and the writers of the transition school (Burke mentions them specifically) are seen as deserters, creating their own worlds which are thought to have nothing to do with the larger world. (Barthelme 1997)
And may even have been influenced by her:
In Tender Buttons Stein either scrambles syntax, or she retains a more or less normal syntax but then fills in the slots with lexically unexpected words…Barthelme seems to be influenced by these possibilities, syntactic and lexical, while using them sparingly, so that they do not overwhelm the reader. (Phillips 1985)
Turing was consumed with the mathematics of machine thinking. But perhaps Barthelme, always on the hunt for interesting language, was consumed with the prose. Back to that opening sentence:
Q: Do you believe that this machine could be helpful in changing the government?
I believe this isn’t two human beings talking about a machine, but one human being asking a machine about itself.
By a mechanical process
“I’m very interested in…sentences that are awkward in a particular way” (Gates 2003)
It’s 1968. Barthelme, let’s say, is contemplating the machine mind. How would it learn? How would it converse?
In my interpretation of this story, the human (Q) offers prompts and the machine (A) “responds.”
Barthelme didn’t know about neural nets (to my knowledge, though they were invented in 1940), couldn’t know that generative AI chatbots would crawl a hundred billion words to guess what word comes next. But his machine is learning. It is “machine learning” — a term that was coined in the 1950s. (So was “Turing Test” — a test to determine if a machine can converse like a human. ChatGPT passed the Turing Test in 2024.) He approximates this using two techniques: robo-talk and repetition.
I think robo-talk is Barthelme getting his feet wet exploring a new form. Repetition comes remarkably close to what ChatGPT is actually doing.
Robo-talk
Q: Is the novel dead?
A: Oh yes. Very much so.
Q: What replaces it?
A: I should think that it is replaced by what existed before it was invented.
Q: That same thing?
A: The same sort of thing.
Q: Is the bicycle dead?
The machine is excited (oh yes!) for the yes/no question because it’s easy to mimic (tangentially, Barthelme wrote about mimetics in Not-Knowing). But the machine breaks when it has to say something concrete. “I should think that it is replaced by what existed before it was invented” is a remarkably convoluted nothing response. It is also a remarkable sentence in the annals of writing.
Here is more robo-talk about a woman ostensibly getting undressed:
Q: Do you see what she is doing?
A: Removing her blouse.
Q: How does she look?
A: …Self-absorbed.
Q: Are you bored with the question-and-answer form?
A: I am bored with it but I realize that it permits many valuable omissions: what kind of day is it, what I’m wearing, what I’m thinking. That’s a very considerable advantage, I would say.
Q: I believe in it.
Early chatbots did talk like this. Here a real conversation between a human and ELIZA, the chatbot developed by Joseph Weizenbaum at MIT in 1966:
ELIZA: Is something troubling you?
You: Men are all alike.
ELIZA: What is the connection, do you suppose?
You: They’re always bugging us about something or other.
ELIZA: Can you think of a specific example?
You: Well, my boyfriend made me come here.
ELIZA: It is important to you that your boyfriend made you come here?
You: He says I’m depressed much of the time.
ELIZA: I am sorry to hear that you are depressed.
You: It’s true. I am unhappy.
ELIZA: Can you explain what made you unhappy?
Barthelme’s machine starts off like ELIZA. But by the end of the story, it has created a fully-formed thought and even referenced its own mind.
Here is another exchange:
Q: She sang and we listened to her.
A: I was speaking to a tourist.
Q: Their chair is here.
A: I knocked at the door; it was shut.
Q: The soldiers marched towards the castle.
A: I had a watch.
Q: He has struck me.
A: I have struck him.
Q: Their chair is here.
A: We shall not cross the river.
Q: The boats are filled with water.
A: His father will strike him.
Q: Filling his pockets with fruit.
Repetition
Large language models (LLMs) work by scanning billions of pages of human written text and generating a ranked list of probabilities for the next best possible word. For the text “The best thing about AI is its ability to,” ChatGPT generated this list of words along with their probabilities (Wolfram 2023):
learn: 4.5%
predict: 3.5%
make: 3.2%
understand: 3.1%
do: 2.9%
The model does this over and over again, picking from a list of best possible next words. There is a parameter called “temperature” (loosely related to randomness) and Wolfram equates “zero temperature” with picking the #1 most probable word at every single step. Here is the result:
The best thing about AI is its ability to learn from experience. It’s not just a matter of learning from experience, it’s learning from the world around you. The AI is a very good example of this. It’s a very good example of how to use AI to improve your life. It’s a very good example of how to use AI to improve your life. The AI is a very good example of how to use AI to improve your life. It’s a very good example of how to use AI to (Wolfram 2023)
So “zero temperature” gives us readable sentences, but it quickly gets caught in a feedback loop. This doesn’t sound like a human being at all. At “temperature 0.8” it is picking the top word but only some of the time (emphasis mine):
The best thing about AI is its ability to automate processes and make decisions quickly and accurately. AI can be used to automate mundane tasks, such as data entry, and can also be used to make complex decisions, such as predicting customer behavior or analyzing large datasets. AI can also be used to improve customer service, as it can quickly and accurately respond to customer inquiries. AI can also be used to improve the accuracy of medical diagnoses and to automate the process of drug delivery. (Wolfram 2023)
And Barthelme:
A: There’s no doubt in my mind that the ballplayers today are the greatest ever. They’re brilliant athletes, extremely well coordinated, tremendous in every department. The ballplayers today are so magnificent that scoring is a relatively simple thing for them.
A: I was standing on the corner waiting for the light to change when I noticed, across the street among the people there waiting for the light to change, an extraordinarily handsome girl who was looking at me. Our eyes met, I looked again she was looking away, the light changed….
Note that “Q” has moments of repetition too, but they are instances of anaphora, a strategic repetition at the beginning of successive clauses to make a point:
Q: Reasons and conclusions exist although they exist elsewhere, not here. Reasons and conclusions are in the air and simple to observe even for those who do not have the leisure to consult or learn to read the publications of the specialized disciplines.
GPT-3’s repetition of “AI can be used” is similar to Barthelme’s cyclic stilted repetition of “the ballplayers today” and “waiting for the light to change” fifty-five years earlier. The syntax essentially predicts a LLM’s “0.8 temperature” response to a prompt.
While I am making this insane case, I’ll briefly mention Northrop Frye, who wrote “probably the single most influential work of literary theory ever written by a North American critic.” (Damrosch 2020) He said that an insane position — okay he didn’t say that, verbatim — can fly as long as it's coming from a real conceptual framework.
In 1983, fifteen years after “The Explanation,” Barthelme writes a story called “Earth Angel.” This is how it starts:
Q: Do we really need Superman III?
A: Clearly not.
Q: Yet it’s here. Must be a response to something, some kind of need…
A: Financial exigencies undiscussable on the plane of the cultural slash aesthetic.
You probably know where I am going with this. In the course of fifteen years, the machine has become much more responsive, going from
Oh yes. Very much so. / I should think that it is replaced by what existed before it was invented.
to
Clearly not / Financial exigencies undiscussable on the plane of the cultural slash aesthetic.
The machine is conversing. It has a real answer to the question (“Financial exigencies”) but still speaks like a machine (“undiscussable on the plane of the cultural slash aesthetic”). It has other tells like fly…in the air (where else do you fly?):
Q: To which we shall stalwartly adhere. Would you like to be able to fly?
A: I’ve always wanted to fly. In the air.
Q: A basic human year. To fly.
By the way, I don’t think that last line is a wistful throwaway comment. I think the human is still dropping hints to the machine about how to be human — saying yes, humans have a desire to fly. In the air.
But aside from these little tells, the machine has gotten quite good. The human’s fifteen years of effort in machine learning are working.
I think it’s only fitting to give ChatGPT — now GPT-5 — the final word.
August 13, 2026.
*At the 2026 ICM, Fields Medalist Terence Tao gave a talk called Mathematics in the age of AI. Here are his slides.
Anderson, David. “Max Newman: Forgotten Man of Early British Computing.” Communications of the ACM. (2013) https://cacm.acm.org/opinion/max-newman/
Barthelme, Donald. Not-Knowing: The Essays and Interviews. Counterpoint. (1997)
Barthelme, David. Sixty Stories. Penguin Classics. (2003)
Gates, David. Introduction. Sixty Stories. Donald Barthelme. Penguin Classics. (2003)
Hodges, Andrew. Alan Turing: The Enigma. Princeton University Press (2014)
Lamb, Evelyn. “Solved, Unsolved and Unsolvable: The Status of Hilbert’s 23 Problems in Mathematics.” Simons Foundation. (2026) https://www.simonsfoundation.org/2026/06/18/solved-unsolved-and-unsolvable-the-status-of-hilberts-23-problems-in-mathematics/ (this isn’t cited above but has some good information about Hilbert’s 23 problems)
Menand, Louis. “Saved from Drowning.” The New Yorker. (2009) https://www.newyorker.com/magazine/2009/02/23/saved-from-drowning
Owen, Chris. “The son of Turing's mentor on beating the codebreaker at Monopoly.” Wired. (2012) https://www.wired.com/story/turing-mentor-max-newman/
Phillips, K. J. “Ladies' Voices in Donald Barthelme's The Dead Father and Gertrude Stein's Dialogues” The International Fiction Review, 12, No. 1 (1985)
Poonen, Bjorn. “Hilbert’s Tenth Problem Over Rings of Number-Theoretic Interest.” These notes form the basis for a series of four lectures at the Arizona Winter School on “Number theory and logic” held March 15–19, 2003 in Tucson, Arizona. The author was supported by a Packard Fellowship. (2003) https://math.mit.edu/~poonen/papers/aws2003.pdf (this isn’t cited either but MIT Distinguished Professor of Science Bjron Poonen writes about what Hilbert’s 10th Problem means today.)
Wallace, David Foster. David Foster Wallace: The Last Interview: and Other Conversations (The Last Interview Series). Melville House (2012)
Wolfram, Stephen. What is ChatGPT Doing…And Why Does It Work? Wolfram Media Inc. (2023)