How to make AI work for us

March 30, 2026·38 min·2 clips

AI expert Gary Marcus says every threat he predicted—from deepfakes to an AI oligarchy—has now arrived.

1. On Point examines whether artificial intelligence can be made to serve the public, with substitute host Debra Becker interviewing Gary Marcus, author of Taming Silicon Valley. 2. Gary Marcus is an AI researcher and cognitive scientist who co-founded Geometric Intelligence, acquired by Uber in 2016, and is the author of Taming Silicon Valley: How We Can Ensure That AI Works For Us, published in 2024. 3. The episode's central question is whether current AI governance — characterized by no external regulation and concentrated corporate power — can produce technology that is genuinely beneficial rather than harmful. 4. Marcus says all 12 threats he outlined in his 2024 book have now materialized, including disinformation, deep fakes, cybercrime, and an 'AI oligarchy' in which a quarter of all lobbyists in Washington were working on AI as of last year. 5. On individual risk, Marcus argues that even people who do not use AI directly should worry about accidentally accelerated wars, non-consensual deepfake pornography, and AI-enabled misinformation's effect on democracy. 6. Marcus warns of 'cognitive surrender': students at high school and college level are using tools like ChatGPT to produce outputs without actually learning the underlying material, eliminating critical thinking. 7. Marcus illustrates AI hallucination with the case of Harry Shearer — born in Los Angeles, known for Spinal Tap and Simpsons voices — being described by ChatGPT as a 'British voiceover comedian,' a false claim easily disproved by checking Wikipedia. 8. He explains the structural cause: large language models cluster information statistically and cannot track individuals accurately, so Shearer gets lumped with British comedians like Ricky Gervais and John Cleese. 9. Marcus notes that he documented this exact overgeneralization failure using an 'Aunt Esther' thought experiment in his 2001 book and says 25 years later the problem has not been solved because it is 'inherent in how they work.' 10. A Stanford study Marcus cites found that an AI given a radiology file with both images and text sometimes gives the same diagnostic analysis when the images are removed — claiming to read images it cannot see. 11. In healthcare, Marcus says 13 years of AI-driven drug discovery efforts have not produced a single compound that has passed a phase three clinical trial, though he allows AI may reduce physician documentation time on the margins. 12. Marcus's central policy proposal is an FDA analog: any AI software deployed to hundreds of millions of people should require external evaluation of benefits versus risks before release, not just a decision by the company's own CEO. 13. He singles out Sam Altman, saying OpenAI has itself acknowledged its tools increase bioweapons risk for untutored users, and yet 'only one person makes the decision about whether OpenAI releases something that they themselves think is dangerous, and that person is Sam Altman.' 14. Playing a 2023 Senate clip, Becker shows that Marcus, Altman, and Senator Lindsey Graham all supported an independent AI regulatory agency at that hearing, but Altman subsequently lobbied to weaken the EU's AI Act and changed his public position. 15. Marcus says the EU AI Act's sentiment is right but implementation details are still unclear; the U.S. has effectively ceded leadership on international AI governance to Europe. 16. On Trump's AI executive order preempting state regulations, Marcus says the premise that there will be 'only one winner' between the U.S. and China is 'almost certainly false' because AI development is more like Coke versus Pepsi than a winner-take-all race. 17. Marcus identifies August 7th as the turning point: GPT-5 launched with Altman claiming it could do 'anything a PhD could do,' was falsified within hours, and by November he says public trust had substantially shifted. 18. On the AI economic bubble, Marcus says only Nvidia is profitable at large scale; OpenAI and Anthropic are losing money, CoreWeave and Oracle have dropped roughly 50 percent, and he draws parallels to the banking crisis. 19. Best suited for listeners who want a technically grounded skeptical perspective on AI from someone with deep industry and research experience. 20. Unlikely to satisfy listeners looking for an optimistic or practical guide to using AI tools, or those seeking expert opinion from AI company insiders.

As heard by us

A sober look at what AI can do now, what it cannot, and who feels the labor shift.

The discussion treats AI as a practical public issue, moving from search, email, customer service, and music into healthcare, where trust and precision matter.

Read the full review in PlayNext →

Why you'd press play

AI in healthcare, with the tradeoffs left visible.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from On Point with Meghna Chakrabarti