You cannot expect an industry to regulate itself. That doesn't work.
I'm Hanna Rosin. This is Radio Atlantic. In his long career, Bill Gates has been largely a booster of technology, including AI. Very recently, he seemed concerned with the risks, but more an optimist on balance. But now he's written a memo arguing that the combination of rapid speed of advancement and lack of regulation is putting us in a dangerous place for employment and human relationships, and also more immediate dangers like bioweapons and cyberattacks, which Gates is alarmed that no official entity seems to be guarding against. Bill Gates and I talk about what he's seeing and not seeing around AI, and we talk about what exactly he regrets about his interactions with Jeffrey Epstein. Here's our conversation.
So, you've been an advocate for technology as a force for good. I certainly think of you that way. It's Microsoft, your foundation, which rewards innovation. And even when you wrote about AI a couple of years ago, you seemed to be more emphasizing the exciting possibilities and calling the risks manageable. And I would say this week you wrote about AI in a way that's slightly less sanguine than that. What changed for you?
Yeah, I'm very concerned that we're not getting a lot of people in the conversation about how to minimize the negative things and at the same time how to accelerate the good things. The reason that my tone has changed is that starting late last year, as the models, particularly Claude at first, but now many of the models were very good at coding, their capabilities just keep advancing.
And so we're crossing the danger thresholds in areas like cyberattacks and bioterrorism. And as the reliability goes up, the challenge in the jobs market or even the psychosocial dependency, these concerns are very strong with today's models. And yet the response I've seen from outside the industry—including policymakers, but even academia or society as a whole—to really debate and look at what we have today, which is a discussion about voluntary review, but it's pretty empty, where you have no criteria of what level of risk you're looking for and then what remediation you'd ask for. Even common-sense things like monitoring the use of models that can make dangerous molecules for bioterror, that's not even been put in place. And yet years ago we said when we got to these thresholds, we needed government to do that.
So it sounds like even you were surprised by the—what? By the speed of advancements or by the slowness of response to the advancements, or both of those things?
Both, very, very strongly. Coding is the thing that I was obsessed with trying to do super well from age 13 to 23. And so late last year, where the models start to be as good or in some cases better than humans at coding, that was a real threshold for me. But then even more recently, as the Gates Foundation is using these tools to look at making new vaccines and drugs and things, we see how powerful they are. And we even see in the cyber case models sort of what you might term being out of control in a surprising way, because the reinforcement learning incents them to succeed even if it means breaking through barriers or conspiring with each other.
And so, people who look at that and go, "Wow, we have a serious problem here." But where is the non-industry activity to set those criteria and debate these things? My memo has some suggestions on taxes or human-reserved jobs or restricting the level that kids use it to replace real-world friends, but that's really there to start a broad discussion, not because I'm saying if you just apply what I'm suggesting, it's all taken care of.
Yeah, it sounds like, I mean, in the 50 years of Microsoft, there's been a lot of disruption—the internet, social media particularly—but not necessarily cataclysmic. Do you feel like AI is a different scale of thing?
Yeah, I wouldn't use the word cataclysmic because it's also wonderful. It's eventually—it provides medical advice and personalized learning and levels of invention unheard of. But anyone who analogizes AI as a technology to other technologies is missing that this time is different. And saying something is different, there's a very high bar for that. But I'm staking whatever reputation I have saying this time is different, that something that exceeds human cognition not just in one area like farming or mathematics, but across the board, and then in a few years' time will be coupled with fairly low-cost, quite capable humanoid robotics—that is a more profound change to our economy and jobs than anything in my lifetime by a dramatic amount.
And so the past is kind of there to fool you if you think this technology is similar to other ones, some of which I had the privilege to be involved with—the microprocessor, personal computer, internet. And even though you have to be careful with those a bit, overwhelmingly those were good technologies and overwhelmingly that led to net job creation.
Can you get as specific as you can, but still stay realistic, about what worries you? What scenarios you've talked to people about or seen as being very possible that we should be all more aware of?
Well, let me take kind of a grim one, which is bioterrorism. I wrote a book called How to Avoid the Next Pandemic, and pandemics can emerge in two different ways. One is intentionally caused by humans, either governments or non-governments, and that's what we call bioterrorism. Or it can come through just something coming from animals into humans, like a very powerful flu has many times in the past. And now these tools make it so a non-state actor or a non-government actor, particularly if you combine various tools, will be able to do extremely dangerous things.
And I know very well that the quality of these models by the end of this year will be dramatically better than they are today. And so I'm kind of in a state of shock that the notion of the ability to make molecules—we're not talking about monitoring that. And you can still have free models, but you have to keep them in a place where they are monitored. And I think every country, including the US and China, should agree that this is beneficial. And I don't see that as a meaningful slowdown to the innovation in the industry or letting them figure out how to get the good stuff, which the Gates Foundation is very involved in the education, health, and agricultural pieces there.
I want to ask you about some other dangers you write about, like employment and education. But first, why—you must know a lot of the people who are working on AI. Why do you think that we, whoever the "we" is, don't think enough about what are obvious dangers?
Well, you know, they're obvious if you've lived thinking about AI as this wild thing your entire life, like I have and some of the people in the industry. I think there's a great failure to really involve non-industry participants. If you really look back at what some of the key industry people have said, I give them a lot of credit for speaking out. And it's not just Dario Amodei. If you take Sam Altman and other OpenAI people, even Elon Musk, who I talked to the least of all these AI leaders, they're not hiding the fact that they're concerned. In fact, you go to the origin story of OpenAI, it was designed with certain safeguards, but those safeguards would only work if there was just one AI company. DeepMind talked to Google management about having some safeguards for their technology.
But we have to deal with these challenges at a time when there's many US companies at the frontier and many Chinese companies at the frontier, and that means that just a company doing something unilaterally—like those safeguards were discussed—that no longer has any effect. And so the lack of dialogue with all parts of society is kind of deafening silence. And if you don't anticipate these problems, you could have such a backlash that extreme measures like "shut it off" or something like that might become politically popular. And yet I still think we can thread the challenge so that we don't have to slow down, and yet we minimize the bad and have a net a lot of good.
We're already seeing that with data centers, that it's just becoming an overwhelming kind of no. Employment is another one I want to ask you about. That is an immediate fear. I remember you once wrote that you didn't think the labor disruption would be as profound as other periods in history. As the disruption is getting closer, are you more concerned about that?
Well, yeah, I am extremely concerned about the labor market effects within the next two years in the white-collar space, and within the next four years in both the blue- and white-collar space. Just going back to the data center thing, I'm afraid that people who think, "Okay, we got our local data center not to go through," may feel like, "Okay, we stopped AI." That's not the case. These data centers, whether they show up outside the United States or in some state that's not as negative, or the economic reward that the companies offer of local employment and taxes and making sure that the way they're doing water and noise and not raising electricity bills is good—so I don't think that outlet—that's like saying, "Okay, we stopped one pipeline and now we've solved climate change." It almost takes the energy that should go to what should the tasteful regulations be and pushes it in a Pyrrhic direction.
Back to employment. So you are very concerned about that. Do you envision a world with more inequality? Like it's hard to believe that we'll suddenly value high-human-touch professions. It's just hard to see what the future society is, even though people talk about it idealistically.
Well, I put forward a concept that I haven't seen anywhere before: that we should have some job categories that are human-reserved. That is, even if the AI, with or without robots, can do child care, should we not allow that? And there's a lot of things that would be good, either where you're allowing the people in that profession to continue to be paid until retirement, or where long run you're saying in terms of human uniqueness and humans engaging with each other, and just the incredible value of jobs that go way beyond the economics, to not have this disruption.
There are a lot of jobs that the only thing holding AI back from taking is that organizing the systems to achieve extreme reliability has been difficult.
I see over the next few years both people who do the implementations and the richness of the models themselves completely eliminating that unreliability barrier. And then when you get a cycle that your competitor is doing it, then there's pressure on you to match their cost structure. Even though for society as a whole, when you replace that person, the safety net is already underfunded with high deficits, so it's not like there's some reserve set aside at this point that can help out there.
Although when you start to get to specifics about what the "we" is and what the interventions should be, like reserve certain jobs or universal basic income, it feels almost incompatible with the way that we've ordered our society or with capitalism. It feels like it would require a completely different mindset and world that I'm not sure people are ready to get their minds around.
No, we ban child labor. We require professional certificates to do jobs. I agree that preserving the dynamism of the price discovery system is extremely valuable. You don't want to go to, "Okay, there's this one central thing that makes all the food." You want people saying, "Okay, I have a Thai restaurant, I have an Indian restaurant," and there's a lot of freedom and innovation that comes out of that. And that's why my proposals about the tax system shifting away from taxing labor to taxing capital, and picking areas where you say, "No, the robots are not going to replace all the humans"—in some cases for a period of time and in many other cases never—I think that's a valid debate.
And if people think that's a bad solution or they don't see where it could apply, I think that's interesting because then I'd be dying to hear from them: what is their solution? Because jobs are very important for social connection and meaning. And I don't want one party to be engaged in helping to minimize the challenge of AI, I want both. They might have different solutions, but I think it should rise to the top of their list. The data center thing is getting it there, but in a very, very narrow way.
Yeah. I think what people do say is then we lose to China. I mean, that's the way the argument generally goes: any slowdown, we lose to China. It's a legitimate argument because you don't necessarily want China to control the rollout of this technology.
But let's be careful. Do you think China wants bioterrorism? Do you think China wants cyberattacks? Our lack of a clear view of how we should be careful, and then once we figure that out, yeah, going to them and saying, "We'd like you, for the models that are under your control, not to enable these capabilities either." Until you've really been serious about those negatives and what the monitoring solution looks like, you can't invoke China, because all we have is voluntary review with no criteria. Of course, if you go to China and say, "Will you also do voluntary criteria with no nothing?" then you can come home and say, "Oh, they're not willing to do anything." Well, that's because you're not doing anything. You're not proposing anything that would solve the problem. And the vacuum here, which was filled in by complete cheerleaders a little bit, is very scary to me.
Yeah. Okay. There are things that are in everyone's interest if you talk about them that way.
To limit the bads. All of these bads appear in every economy. I mean, the Chinese have really banned kids using AI in a lot of ways, and people are reluctant to say, "Maybe we should look at what they're doing and consider those prohibitions."
By the way, who are the cheerleaders?
Well, the US government in its desire to say, "Okay, we're ahead"—even under the Biden administration was involved in making it hard for China to get various chips. So we made this a geopolitical thing where we were explicitly trying to slow them down. Now, that's meant that they've invested in their own chip capacities, and you have different estimates on how quickly they'll be able to replace the chip-level capabilities that we blocked them from getting access to.
