Most People Think They Can Spot AI Fakes and Are Wrong; Rosenbaum Gives the Skill Six Months

Step by step

1. The real problem is confidence, not just fakes ▶5:29

   Nick says several universities have run studies that mix AI-made faces in with real photos of people. People sort them only slightly better than random guessing. Rosenbaum's central worry builds on that. Most people are certain they can tell real from fake, and most of them cannot. He says anyone who can spot fakes today should assume that in six months the skill will be gone. He adds that checking one disputed photo took him a morning. A New York Post article called an image from the correspondents' dinner shooting fake. No other news outlet backed it up, and his judgment was that the Post was wrong. Most readers, he says, will not do that work. They settle for "it could be fake."

2. Video fakes now fool the people in them ▶11:11

   The talk turns to Sora, which could place a person in his own studio with a voice no one could tell from his own. Rosenbaum uses this as proof that judging by eye or ear no longer works. He reads Sam Altman's choice to pull Sora 2 as a business call, not an ethical one. There was no money in fake video, and it put the company at legal risk. He says it solves nothing, because plenty of other sites make convincing fake video. If someone makes a fake of you, Rosenbaum says a complaint to TikTok comes back as "section 230, free speech, tough luck."

3. Cheap AI will not stay cheap ▶24:54

   Rosenbaum predicts a "cost apocalypse" for AI. Today's tools are heavily subsidized, he says, and the companies will start charging a premium. His model is Uber in New York. Early rides cost $9 when a cab was $35 or $22. He says a recent Uber from his downtown apartment cost $57. In his view, the same venture firms that funded that pattern are funding AI. He also worries about paid answers. He recalls Google Maps showing dollar-store logos on a drive upstate, and he ended up at one getting paper plates. With AI, he says, nothing may tell you an answer was paid for.

4. Will people pay more for human-made work ▶33:13

   Rosenbaum raises a hard question against his own hope. If a film entertains him and he has a nice night out with his wife, he is not sure a "human actors" label changes anything. Shoppers rarely ask who picked their food. He expects cheaper and more efficient to feel appealing, and sloppy to feel unappealing. He also points to an age gap in writing. He has written three books and blogs every week, so AI can polish his drafts in his own voice. A 17-year-old has no body of work yet, so the shortcut arrives before the voice is built.

5. Why he is still an optimist ▶37:37

   Despite his worries, Rosenbaum believes the flood of synthetic content will make people value human-made work more, not less. He runs the Sustainable Media Center, which has a board of Gen Z leaders. He says more and more of that generation is putting the phone down and gathering in person. He compares it to early COVID, when life went fully digital and he found it efficient at first, then missed people deeply. His fear is that AI will offer to handle every task for us, and by the time we notice what we gave up, it will be too late.

6. Chatbots state things instead of checking them ▶45:10

   Rosenbaum describes finding an AI-made video laying out his whole career, which he uses as evidence that models assert rather than verify. He says correcting a chatbot does not make it smarter, because your fix is not fed back into its data. Nick adds that a model once gave him nine papers, and two were completely invented but looked just like the real ones. The cost to people is already concrete, Rosenbaum says. A publisher can weigh replacing an editor at a salary of $64,000 with a tool he pays $20 a month for.

7. A model built to tell the truth helps more ▶56:12

   Nick shares an example from his own company, Test Prep Gurus. Students would say a general chatbot told them they were a great fit for Stanford. Nick says Stanford's acceptance rate is three percent, and once donor families and recruited athletes are counted, it is probably less than 1% for a normal kid. So the team built a separate model just for admissions odds and had to wrestle it into being honest. Nick says it has been much more useful to families than the general models. Rosenbaum asks whether people really want the honest version.

8. The future of truth is human ▶80:54

   Rosenbaum closes with his thesis. AI is good at looking backward, like summing up something that already exists. It is poor at looking forward and proposing a future. He tested this by asking ChatGPT to write the last chapter of his book from six elements. What came back repeated earlier material and did not make sense. He does not expect laws to fix things, because by the time a bill passes, he says, ChatGPT "is going to be on version 12." His hope is that society comes to see some AI profits as tainted. The robots can help, he says, but people must keep doing the thinking.