Pangram: Can we trust AI detectors?

AI is causing “a crisis of trust” in publishing, said Vauhini Vara in The Atlantic. Earlier this month, Zimbabwean writer Simbarashe Steyn Kundizeza won the prestigious Caine Prize for African Writing—worth $13,000—for a short story that “has all the personality of a chatbot.” That may not be coincidental: Pangram, an artificial-intelligence-powered AI detection tool, flagged “almost the entire story as AI-generated.” Kundizeza is not the first award recipient to be flagged by Pangram; it deemed one of the winning stories in this year’s Commonwealth Short Story Prize to have been entirely authored by AI. Kundizeza says his story, “The Blue Fighter,” is “entirely my work.” Pangram is reputed to be “highly accurate”—the company claims its tool misidentifies AI use in just 1 out of 24,000 cases—but it’s not infallible. So who should we trust? The author or Pangram?

This is exactly the “chaos” that publishers feared, said Anna Silman in The Wall Street Journal. Publishing houses have tried to crack down on AI usage by authors, inserting strict clauses into book deals and leaning on Pangram to vet manuscripts. But “a new AI scandal engulfing” the industry seemingly every month is “forcing a reckoning over the nature of authorship.” The Big Five publishing houses—Simon & Schuster, Penguin Random House, Hachette, HarperCollins, and Macmillan—have been reluctant to “make sweeping pronouncements about” the technology, possibly because they’re waiting to see if people care whether the books they’re reading are AI-generated or AI-assisted, “especially in genres like fantasy and sci-fi.” Meanwhile, many writers feel that a contract clause “that relies on the honor system lacks the teeth necessary to combat” intellectual dishonesty.

I write a lot, and I use AI, and I’m worried that “I’ve started to internalize” how a chatbot thinks, said Alex Kirshner in Slate. I’ve begun to “self-police against this,” modifying my drafts “to sound less like AI.” I also put my work into Pangram “before it appears anywhere else” to ensure that “the model concurs that I wrote it myself.” It feels dystopian—I’ve had to “rewrite sentences of my own work in hopes of satisfying this new integrity god.”

The Week

Escape your echo chamber. Get the facts behind the news, plus analysis from multiple perspectives.

SUBSCRIBE & SAVE

Sign up for The Week's Free Newsletters

Join more than 350,000 subscribers and keep yourself informed with a selection of The Week’s most interesting, enlightening and entertaining stories - plus daily puzzles.

Join more than 350,000 subscribers and keep yourself informed with a selection of The Week’s most interesting, enlightening and entertaining stories - plus daily puzzles.

Sign up

Still, we need the “AI slop detectors,” said Brian X. Chen in The New York Times. The tools do more than pick out obvious indicators of AI authorship, such as the excessive use of em dashes and bullet points. They’re trained to understand how AI models make decisions on word selection and phrasing. Pangram, which costs $20 a month, is incredibly “useful for sniffing out annoyances like AI-written email scams and phony online reviews.” Yes, the technology “still has a long way to go.” But “at the rate that slop is proliferating on the web,” it’s an “empowering” tool for people who still want to read the human-written word.