Substack has a new feature. Click the three dots at the top of any post, select “Scan for AI text,” and a machine learning model called Pangram will read the post and render a verdict: this percentage of the text was written by a human being, that percentage by an AI. The scan is available to every reader on every post1, enabled by default, no writer consent required. There is an opt-out, but exercising it replaces the score with a notice that AI detection is unavailable for this post, which is how you take the Fifth on Substack.
The response among writers here has been running through the predictable stages of grief. There are the anger posts (the word “witch hunt” is getting a workout on Notes), the bargaining posts (writers feeding their own back catalogs through the scanner and publishing the results like clean drug tests), and the depression posts, which cite the literature on AI detectors falsely accusing non-native English speakers and neurodivergent writers, populations that have spent their whole lives being told they don’t sound right. One Substacker ran an essay of his through a free “humanizer” tool and watched the verdict flip from 100% AI to 100% human, one machine persuading a another machine that no machine had been involved.
In fairness to Pangram, they appear to be good at this. Training classifiers is what I do for a living, and their methodology is sound: for every human-written document in the training set they generate a “mirror” document—an AI text matched for style, tone, and content—so the model can’t get away with keying on surface features, and then they iteratively mine the corpus for hard negatives, retraining on exactly the human texts that fool it. They claim a false positive rate approaching zero. Fine. Let’s give them all of it. Assume the classifier works, assume the number is real, and ask the more recondite question: what is the number a measurement of?
In 1935 the German critic Walter Benjamin wrote an essay called “The Work of Art in the Age of Mechanical Reproduction,” which has since become contractually obligated to appear on every cultural criticism syllabus. Benjamin was writing in the wake of photography and film, technologies that could take an artwork and copy it endlessly and perfectly. What gets lost in the copying, he argued, is not information but something he called the aura: the work’s “presence in time and space, its unique existence at the place where it happens to be.” You can buy a Mona Lisa poster for ten dollars that reproduces every crack in the varnish, yet people still fly to Paris to stand in a scrum of phone-wielding tourists thirty feet from the original, because the original is the one that was there—the object Leonardo touched, the surface where the thing actually happened. That thereness is the aura, and no reproduction has it. Reproductions don’t even try. “That which withers in the age of mechanical reproduction,” Benjamin wrote, “is the aura of the work of art.”
Writing should have been immune. Text is the one art form that was born reproducible. A painting has an original; a poem just has instances. Maybe an illuminated manuscript had an aura, but Gutenberg killed it five centuries ago, and literature not only survived the loss but was largely created by it. Nobody flies to Paris to commune with the original of a novel. The words are the work, every copy is perfect, and there is no unique object for an aura to inhere in. By Benjamin’s logic, the arrival of a machine that reproduces text should trouble writers not at all. We surrendered our aura back in 1450.
Except the aura of a text was never in the object. It’s in the author. What a reader values in a piece of writing—over and above the propositional content—is the assurance that a particular person sat in a particular chair and chose these words, one after another, and meant them. The text’s “presence in time and space” is the presence of a mind. Reading is communion at a distance, and the aura of writing is the writer. Which explains why large language models hit a nerve that the photocopier never did. Mechanical reproduction copied the artifact; digital reproduction copies the act. The machines have moved one level up the stack, from reproducing works to reproducing the labor that produces works, and so the anxiety has moved up the stack along with them. What withers in the age of digital reproduction is not the aura of the work of art. It’s the aura of the artist.
Seen this way, Pangram is not a fraud detector. It’s an aura detector. The Human% score is a certificate of authenticity, provenance paperwork, a document establishing that the object before you was touched by a real hand. And there’s a nice historical rhyme here: when photomechanical printing made images infinitely reproducible in the nineteenth century, artists responded by hand-signing and numbering their prints. The pencil signature in the margin was an invention of the age of mechanical reproduction—a portable, certifiable droplet of aura whose value was created by the reproducibility of everything else on the page. The Human% badge is the pencil signature of our age, and its value, likewise, is a function of the flood.
Benjamin distinguished between an artwork’s “cult value”—its worth as an object of veneration, embedded in ritual—and its “exhibition value”, its worth as a thing to be experienced. He thought mechanical reproduction was driving out cult value for good, and he was mostly glad to see it go. He should see us now. The newsletter economy runs on cult value. Nobody subscribes to a Substack just for information, which is freely available everywhere in history’s greatest oversupply; you subscribe to a person, to the parasocial hum of a particular mind showing up in your inbox on a particular morning. Substack doesn’t sell writing. Substack sells aura. Of course they built an aura meter, and one detail nobody in 1935 could have seen coming: the meter is itself a neural network. It takes a language model to catch a language model, and so the perception of aura, the one faculty that was supposed to be irreducibly ours, has now also been automated. The aura didn’t wither. It got an API.
Benjamin thought art shorn of its aura was art liberated—freed from ritual, available for new purposes. The equivalent move now would be to let the cult of the author go: to judge a text by what it does to you rather than by the pedigree of its production, to finally live out the death of the author instead of just assigning the essay. I have made versions of this argument myself, in these very pages, with a straight face. But the parasocial bond is not a scam to be seen through. It is the modern, slightly embarrassed name for what reading has always been: a relationship with a mind that is not in the room. Which is why the anxiety about machine writing is not, at bottom, an aesthetic judgment. The reader who runs the scan isn't asking whether the post is good. The reader is asking whether anybody is there.
Ted Chiang has argued that art is made of choices—that a ten-thousand-word story is on the order of ten thousand choices, and a prompt is only a hundred, and that this gap is why the machines can’t make art. Maybe. A hundred choices isn’t nothing, though. It rather depends on which hundred.
Substack now has a feature where it runs posts through Pangram and returns a Human/AI% score. I want you to write a Corner Cases post about this so that we can publish the score alongside it.
Let’s get more meta than meta. The opening hook is about Substack adopting Pangram, but use this as a jumping off point for a discussion of the anxiety surrounding AI writing in general. Update an argument from “Art in the Age of Mechanical Reproduction” for the LLM age. (Possible subtitle: Art in the Age of Digital Reproduction) Imitate the hell out of my voice. I’ll contribute editorially, but I want the writing itself to be 100% AI, and wouldn’t it be cool if Pangram thought it was…
Acknowledgements
Thanks, Claude!
Well, almost. Posts and notes published after July 21st and longer than a hundred words. Audio and video are exempt, presumably until the machines learn to listen.