The essay that read like filler, and what that says about prompts

Stephen Marche's piece in The Atlantic opens with a test. Suppose you're a professor and a student hands in an essay on learning styles. He prints the first paragraph. Then he tells you it was written by GPT-3, prompted by Mike Sharples, a professor in the UK, who considered the whole essay graduate-level.

Marche's own grade: "Personally, I lean toward a B+. The passage reads like filler, but so do most student essays."

The piece was titled "The College Essay Is Dead" and went up at 11:30 AM ET on December 6, 2022.

The argument

Marche, a former Shakespeare professor by his own description in the piece, mentioned that OpenAI had released ChatGPT the week before and that it had set off a new wave of marveling and hand-wringing. He quoted University of Toronto associate professor Kevin Bryan, who had tweeted that you could no longer give take-home exams or homework.

Most of the essay, though, is about two groups that don't talk to each other. Cheating is only the way in. Marche leaned on C. P. Snow's old "Two Cultures" lecture to argue that engineers had ignored the humanities and the humanities had ignored technology, and that language models would force them together. He guessed it would take academia about ten years to face the change: two for students to figure out the tech, three more for professors to notice, five for administrators to decide what to do.

You can agree or not. I think he was too pessimistic about the timeline and right about the gap.

The line I keep coming back to

"Reads like filler." That's the most honest critique of AI writing anyone made that week, and it's still mostly true when people prompt the way most of us do.

Filler comes from a prompt with no material in it. Ask for an essay on learning styles and you get the average of everything ever written on learning styles. It's smooth, it's correct-ish, and it says nothing a reader will remember. The model isn't lazy. It's doing exactly what you asked, which was to write about a topic without telling it what you think.

The fix is to bring your own stuff. A position. Three facts you found. An example from your life or your work. When the model has to build from your material, the result can't be generic, because your material isn't.