“You can't have it both ways, Brian. The tech industry tells us that AI is the most important technology since fire. It will shape every aspect of our lives. It will completely transform everything we do. But then they're also telling us, stay out and let this be decided by a handful of people.”
Into the Impossible With Brian Keating · Ideas & Essays · October 2026
Acemoglu had been arguing that whether AI replaces workers or gives them new capabilities is a design decision, and that the industry is currently betting everything on replacement. His conclusion was that a technology this consequential has to go through democratic debate and regulation.
Transcript
It should be.
Okay, so explain that argument. Yeah. How could it be there? What example? Use teaching. Use teaching as an
example. I am very worried about job losses from AI. I'm the first one, and the paper that you mentioned emphasizes that. I'm the first one to stress that we're not seeing just job losses yet. It's not something like it's going to happen by the end of 2026. But potential job losses from AI, if it goes down the AGI path and tries to automate everything, would be quite disruptive, both because they would create huge inequality, but also politically and socially. And my argument is that that's not the only path for AI. We can also use AI to create new tasks, new capabilities, new expertise for workers. And that would be both productivity enhancing and socially more beneficial. But I'm not saying that that's necessarily going to happen because whether we do that or not is a design decision. And right now, we're putting all of our design eggs in the AGI basket. So this is why I think we need a framework for the regulation of AI and support for AI from government entities. But even more importantly, we need a democratic debate on AI. You can't have it both ways, Brian. The tech industry tells us that AI is the most important technology since fire. It will shape every aspect of our lives. It will completely transform everything we do. But then they're also telling us, stay out and let this be decided by a handful of people. That can't be, if it's half as important, a quarter as important as they say. It's the most important thing that has happened to us in the last 100 years. And therefore, of course, the democratic process must have a say in how we develop AI, how we use it, what we do with it, who benefits from it, who pays the cost. That's a democratic decision. We are still a democracy, perhaps not for long, but we are still a democracy. And the reason why we're not guaranteed to be a democracy in a couple of years is because this is why I wrote the book, Liberal Democracies in Crisis, and AI is deepening that crisis.
Today you tweeted about the alignment problem, which I've talked a lot about with people from Jan Lacun to Max Tegmark to Nobel laureates as well. But the question that you raise is that the problem that people are mistaking is not the capabilities, as you say. The capabilities are real. But more than the capabilities is, I quote, some imperfect quantitative metrics, user approval, user engagement, simple completion metrics. This is the process that leads to distorted behaviors in the form of gaming the evaluation of completion metrics, cheating, overconfidence and wrong answers, sycophancy. Although I love sycophancy, Durham, because I asked AI once, I said, what books has Brian Keating written? And he said, Losing the Nobel Prize into the Impossible and A Brief History of Time. I thought that was great. If I can get 1% of Stephen Hawking's book sale. Okay. And then you say an analogy may help. It's not that we have a supercar that has its own mind and wants to take control of driving because it has superiority to the driver. It's more that we have a car where the steering and brake systems don't work. It has many capabilities of very good cars, and its engine accelerations and GPU may be very impressive, but you cannot steer it properly. What good is it? Perhaps we shouldn't drive it until it's fixed. How would you fix it? You have control. What do you do to align it the way that you would best see fit?
Look, I refuse that control. I'm not an expert on AI. But my assessment is. That there are so many good things we can do with domain-specific models. You want to use AI in the production process, but better quality control, better design, better training, make the workers more productive. All of these are domain-specific problems. AlphaFold is a domain-specific AI model. Those models don't behave like cars without steering because they are designed for something very simple, very specific. Because they are designed for something specific, their metrics are much more aligned with what we want. It's these general models that are trained on all the literature, all the crazy stuff that's ever said on Reddit and Twitter. And then we want them to have as much agency and decision rights as humans. That's where the problem lies. So I think if we train these models much more for specific purposes and deploy them for specific purposes, we're going to do so much better.
All right, Duron, I know you got to go, but I have one more question for you. Please. I want to leave you with a line that I think is quite beautiful. I'm sure you've known it your whole life because it comes from Ataturk. He said, For civilization, for life and for success, the truest guide is science. You spent your whole career studying institutions, working at, arguably, outside of UC San Diego, the best scientific institution in America.
More from Into the Impossible With Brian Keating
