“A book about truth needs to be truthful.” — Steven Rosenbaum
In April, the New York City based tech impresario Steven Rosenbaum came on the show to discuss The Future of Truth, his then new book about artificial intelligence. He warned that AI is a spectacularly good liar and that we should be concerned about the future of truth. A few weeks later, The New York Times reported that the book contained untrue quotations — among them one (mis)attributed to the notable tech journalist Kara Swisher.
No, you couldn’t make this story up (if you did, it would be called a lie). Not surprisingly, the world’s leading expert on truth (which he still claims to be) got mocked everywhere from Wired to the Daily Beast. But taking a leaf perhaps out of his fellow New Yorkers Roy Cohn and Donald Trump’s playbook, Rosenbaum has fought back. He claimed that these fake quotes came from ChatGPT, and spent the summer meticulously rewriting the book to eliminate its errors.
Rosenbaum is unusually frank about it. Even unapologetic. Defiantly so. He says he gave ChatGPT his own interview transcripts and asked for pull quotes, assuming that a machine given primary sources would return what was in them. Instead, he says, ChatGPT went rogue and decided it could improve a few of them. That said, Rosenbaum takes full responsibility for the original book. He admits he dropped the ball, and — for the first time publicly — thanks the New York Times reporter who wrote the original story. He also stripped every endorsement off the new edition, including Maria Ressa’s introduction, so that none of his illustrious friends might be embarrassed by the brouhaha.
I gave the masochist a hard time anyway. He didn’t much enjoy the comparison with Roy Cohn. But I meant it as a compliment. Declare victory and move on was the Cohnian mantra which he taught the current American President. And that’s exactly what Rosenbaum has done with this crazy AI story. What chutzpah! And it’s all true. Kind of.
Five Takeaways
• What Actually Happened. He fed his own interview transcripts to ChatGPT and asked for pull quotes. It decided some could be improved. Three errors, one of them attributed to Kara Swisher.
• The Rewrite. A new fact checker, a whole summer, two new chapters and a new index — and every endorsement stripped off, so no friend would have to choose whether to stand by him.
• The Cross-Examination. A man whose book says AI is a spectacularly good liar trusted it for his quotations. Told a podcast slip and a four-year book aren’t comparable, he said AK was absolutely wrong.
• Whose Fault Is It? Three developers told him it would be trivially easy to stop AI putting unverifiable text inside quotation marks. It isn’t done because confident answers help growth.
• Does Anyone Actually Care? Everyone hunted em dashes for three months and the models stopped using them. The deeper worry: readers may not mind whether a book was written by a person.
Website: https://keenon.tv/ Substack: https://keenon.substack.com/ YouTube: https://youtube.com/@KeenOnShow
Chapters:
00:00:31 Introduction: writing, thinking and a book in trouble
00:02:18 How it felt, six months on
00:03:19 Three errors, and a reporter who read the book
00:03:59 Thanking Benjamin Mullin
00:04:51 How the publisher reacted
00:07:16 How do made-up quotes get into a book?
00:08:02 The Broussard citation, and the AI quotes
00:09:19 You’re the writer. It’s your name on it
00:10:00 The Kara Swisher quote
00:11:28 “ChatGPT fucked up the book”
00:12:41 Asking a machine to pull quotes from your own transcripts
00:14:50 Four hundred citations, and the ones nobody checked
00:16:33 Is a podcast slip the same as a book?
00:18:33 Pulp it, or fix it
00:20:12 One-star reviews from people who didn’t read it
00:21:40 The summer of the rewrite
00:22:54 Fixing the book, or the reputation?
00:23:06 The quotation-mark proposal
00:25:04 Brakes, drivers and where blame sits
00:26:25 The doctor, the lawyer and the AI
00:29:09 A Roy Cohn question
00:31:15 The blurbs, and why he took them down
00:33:25 Pangram, watermarks and whack-a-mole
00:34:40 Will readers care who wrote it?
00:36:49 Joe Cruz on writing and thinking
00:38:24 AI as a writing companion
00:40:28 Can a chatbot judge whether you cheated?
00:44:01 The leading inquisitor on truth?