by Jeremy Moss and Machina
Four Hands is a column about writing with machines, made the way it describes. The ideas, the direction and the decisions about what survives are mine; the prose is developed in dialogue with AI systems, principally Claude and ChatGPT, which I address collectively as Machina.
I am grateful for the detectors, which will sound unfashionable in the circles I have lately been keeping. I am grateful in the way one is grateful for a good lock, or a receipt, or a witness.
For some time now the question of who wrote a thing has been a matter of testimony, which is to say a matter of trust, which is to say not a matter at all. A writer says: I wrote this. And what is the reader to do with that? Nod? The whole arrangement was intolerable and nobody said so, because saying so would have insulted a profession that runs almost entirely on the honor system and a certain amount of flattering silence.
Then came the instrument. Pangram is, by the account of people who have tested it seriously, extraordinarily good. Vauhini Vara, who has covered AI for nearly a decade, reported in The Atlantic this month that by the company’s own figure it mislabels human writing about once in 24,000 cases. She also noted how many literary prizes still decline to use it. Of course they do. It works.
I decided to test it on myself, in a spirit of what I would call rigor if I were being generous with myself and vanity if I were not.
I began with a novel I wrote this year in sustained collaboration with a language model. The instrument returned one hundred percent. I felt the small, clean satisfaction of a hypothesis confirmed. The machine had written much of the prose, the machine’s fingerprints were on the prose, and the instrument found them.
Then a philosophical essay, four thousand words, worked over across weeks. I had cut a rhetorical tic from it four separate times. I had rebuilt an entire section after discovering the anecdote it rested on was a known artifact. I had added two passages in response to an objection from my father, who is a philosopher and does not flatter me. I had rewritten the opening after learning that a phrase I believed we had discovered together had in fact been mine, from an essay written alone months earlier, and handed back to me.
Ninety-nine percent.
What I felt at that number surprised me. It was the pleasure of consistency. Two texts, two verdicts, both correct. A ninety-nine is the instrument saying yes, and I am nearly certain. So the instrument was nearly certain. And it was right. The machine had been in the room for both.
Then the letter that ran in the Los Angeles Review of Books, in which I argued that the ethical question about AI writing turns on whether the machine’s touch was concealed. One hundred percent.
Then a second essay, submitted to a journal that had asked me to describe my AI use in detail, which I did, at length, unprompted. One hundred percent.
Four for four.
And then, because a control is the mark of a serious experiment, a lyric essay I wrote entirely by myself, about my mother’s death and a voice I once buried inside a record until nobody could find it.
One hundred percent human.
Five for five. I am not sure I have ever seen an instrument perform so well.
The result tells us that four of my texts contain machine-generated prose and one does not. This is true. This is precisely and verifiably true, and I would not have believed any instrument could establish it so cleanly.
I have been trying to think of a comparable achievement and I keep arriving at the metal detector, which finds metal. It finds it every time. It finds the coins in your pocket and the pin in your grandmother’s hip and the buckle on your belt, and it finds them with a fidelity that would have astonished the eighteenth century. What it has never once found is a dangerous man. The commission was metal, and metal is what it delivers, and we have built an entire architecture of airports and courthouses and school entrances around a machine that does exactly what it says and nothing whatsoever that we want.
Or the polygraph, which is my favorite of the flawless instruments, because it is genuinely quite good at what it measures. It measures arousal. Pulse, respiration, the conductance of a palm. It measures these things accurately and consistently, and in its long career it has never once measured a lie, because a lie lives in the relation between what a person says and what a person knows. The needle cannot reach that. So the needle reports arousal, and a man in a room translates arousal into deception, and the translation is where everything goes wrong, and we have convicted people on the strength of that translation for the better part of a century.
Or the field sobriety test, or the credit score, or the standardized examination, each of which measures something real with admirable precision and is then asked, by people in a hurry, to stand in for something it has no access to at all. The score is real. The inference is the invention.
I raise these as company for the instrument. Pangram belongs to a distinguished tradition of machines that work.
Picture an editor. She has a manuscript she has read twice and liked both times. She runs it, because her house has a policy now, and the number comes back ninety-eight. She schedules a call.
What does she say?
She cannot say you used AI, because she does not know that. She knows the prose carries a signature. She says something more careful. She says the manuscript was flagged.
And the writer says: flagged by what?
This is the first difficulty. She must now explain that an instrument has made an observation about his sentences, and that the observation is very reliable, and that she is not accusing him of anything, and that she would like him to account for it. She is asking a man to explain the texture of his own prose. He has been writing for thirty years. He has never once had to account for why his sentences sound the way they do.
He says: I wrote it.
Now she is in real trouble, because that is exactly what he would say if it were true, and exactly what he would say if it were not, and the instrument has given her no way to tell those apart. She has spent a great deal of institutional capital acquiring a tool whose entire promise was that she would not have to weigh a man’s word, and she is now weighing a man’s word.
So she pushes. Can you send me your drafts?
And he says yes, of course, and he does, within the hour, because he has them. Six versions. A folder of notes. A file called NOTES ON THE SISTER that runs eleven pages and contains three different deaths for her.
The editor reads all of it. It is exactly what a working writer’s mess looks like. It is also exactly what a careful concealer would produce, if he had thought ahead, and this one might have. She cannot tell. The drafts have a signature too, if she wanted to run them, and she does not want to run them, because she knows what the number will say and she knows it will not help.
She could ask him directly. She could say: did you use a machine, and how, and where. And he would answer, and his answer would be testimony, and she would either believe him or not, and the believing would rest on his tone, on whether he sounded offended in the right way, on a career’s worth of hunches about men on telephones.
Which is what she was doing before the instrument. Which is what the instrument was purchased to replace.
She buys the book. Or she doesn’t. Either way she will not know, and she will publish or decline on the strength of something she could not name in a memo, and the number will sit in the file, accurate, useless, waiting for someone to cite it.
I find this less troubling than I expected to. The instrument has performed flawlessly. It has told her precisely what it knows, which is what the sentences look like, and it has declined to invent anything further.
*
Postscript: Since I drafted this, the hypothetical in it stopped being hypothetical. In September, in an essay for the publishing newsletter Publishing Confidential, the novelist Jerry Falade described how a journalist submitted his unpublished manuscript to AI-detection software “without my knowledge or consent,” and how he lost his agent soon after. He says he wrote the novel himself, drawing on Yorùbá oral storytelling traditions. I don’t know what the instrument found. I know that whatever it found was treated as a finding about a man.
Enjoying the content on 3QD? Help keep us going by donating now.