Bill Gates says it’s ‘not enough to have a kill switch’ for AI: Full interview

[0:00] Mr. Gates, welcome back to Meet the Press. Great to be here. It's wonderful to have you. We really appreciate your taking the time. It's a busy week. I want to start off by talking about the initiative that you have announced this week. It aims to address one of the biggest concerns about AI, which is the equity in the AI space, the language barrier AI mostly built for English speakers. And so the Gates Foundation is partnering, if I'm right, with 60 other organizations, including [0:30] AI companies to expand language accessibility. Why do you think this is an important initiative? Why now? AI models have gotten very good at recognizing English and about a dozen other mainstream languages. And so as you talk about your health, it's going to accurately record that and be able to help you. In African languages, the error rate will be about 10 times higher. [1:01] Oh, wow. And so that's very dangerous. You know, take Uribe, which is 3 million people in Western Africa. You have 10 times the error rate that you have in English. So it's really not usable. And yet, you know, pure capitalistic incentives would not prioritize these languages. [1:27] And so it was amazing. And so it was amazing. You know, we hope to get 10 companies working with us. We ended up with 60. And so we have a very concrete timeline to improve the quality, to benchmark and then say, okay, this is good enough. You can do a health interview or an education session in this language. [1:49] What does the timeline look like? What does the timeline look like? What does success look like in your eyes? [1:54] Well, Lee, over the five years, we want to take every language that a million or more people speak [2:01] and have it be close to English language quality. And so you could say that a patient or a farmer or a student, [2:12] it makes sense for them to use that audio interface, you know, talking about their crops and the challenges they face. [2:23] And so, you know, this is extremely well-defined, how we measure the quality, what data we need. [2:32] And all of the model providers will then take the data. And so these capabilities will show up in every single AI model. [2:40] Have you been surprised by how quickly AI has been able to grasp language capabilities in the first place? [2:49] Is that one of the aspects of it that's been surprising as someone who's worked in this space? [2:55] Well, certainly when you go back to the, when ChatGPT came along, I had challenged the people at OpenAI to take a test, [3:06] like an advanced placement biology test. And it was one of the more amazing demos ever in my life [3:15] when we saw that the, you know, so six months before it was released, we saw how good it was at language. [3:22] Language is actually a perfect application for the architecture we use here. And so, you know, [3:30] the idea that all the world's languages can be supported is actually pretty straightforward. [3:36] Does that give you optimism that you can, in fact, meet the benchmarks that you've laid out for these five years? [3:41] Yeah, of all the problems in AI, this one, it's clear what we need to do. It's clear how we do it. [3:49] It's clear, you know, it's over three billion people whose languages are currently very, very poorly represented. [3:59] So almost half the world is getting far, far worse quality than we experience. [4:06] So let's talk about this moment in AI development. You are one of the voices who's been warning about the potential risks. [4:17] You wrote in August, quote, AI will either be the greatest equalizer ever invented or the worst source of injustice. [4:27] How serious are the risks posed by AI, do you believe? [4:31] Well, AI is not just a new technology. And that's why people who say, okay, past technologies, there was all this worry. [4:39] And, you know, by and large, it all worked out. So maybe this time, all these warnings that we have to shape it intentionally are unnecessary. [4:51] This is truly an evolutionary event. You know, we have created an intelligent species that, in many respects, is already smarter than we are. [5:02] And over the next several years, at kind of exponentially improving level, we'll be insanely smarter than we are. [5:12] You know, the ability to transfer knowledge. There's no sort of individual boundaries like humans having to educate each other. [5:20] And so I can't say strongly enough how much society, not just the companies, but every academic, think tank, voter, policy maker, need to really engage on how we want to shape this technology. [5:38] You know, there's no such thing as being too focused on accelerating the good potential and minimizing some truly horrific, bad things that could happen. [5:57] You say something that is truly noteworthy, that in some ways, AI is already smarter than we are. [6:04] Is it possible that AI is already out of human control? [6:09] No. [6:09] Okay. Humans still have control of AI? [6:11] We do. [6:12] What gives you confidence? [6:14] Well, the AIs, they're not running millions of robots yet. [6:20] You know, so their ability, you know, we can turn off the computer. [6:27] And the most dangerous thing we're facing right now is people with bad intent using AI. [6:39] You know, so somebody who wants to defraud an individual or wants to shut down the electric network or, you know, scramble bank accounts. [6:50] They are now empowered. [6:53] Historically, countries, you know, had cyber attack teams and they could think of things like bioterror weapons. [7:03] Now, small groups with AI can do what only the biggest countries could do. [7:10] So it's really the bad intent plus AI that looms as a major threat unless governments insist on various safeguards. [7:27] You know, we will see a lot of big negative things because we're not safeguarding. [7:34] The big headline in recent weeks as it relates to AI was a whistleblower, as you know, within the industry, sounding the alarm, warning in an online post. [7:46] I'll read a part of it. [7:47] The people building AI earnestly believe that it could kill us all by the end of the decade. [7:54] Do you believe AI is powerful enough to end all of humanity? [7:58] AI is certainly powerful enough to drive events that, you know, cause a billion deaths, you know. [8:09] So even though it's pretty hard to get to 100%, there's never been a weapon as powerful as the combination of people with ill intent using the latest AI tools. [8:26] And it's great that there's this discussion with the industry, almost against their interest, admitting these things. [8:34] The problem is that the rest of the world is either discounting this or just thinking of this as a competition. [8:44] And it's very scary that we don't have, you know, all of the thinkers saying, how do we want to shape this? [8:54] So how should people be thinking about this warning from the AI whistleblower who says that AI is strong enough to destroy all of humanity? [9:05] What should be done in this moment? [9:08] Well, certainly you've got to deal with the near-term risk, which is that these models enable extremely strong cyber attacks, extremely strong defrauding individuals. [9:26] And so AI, instead of just running, you have to build the safeguards in, including monitoring, where you're watching what the AI is doing. [9:38] You know, law enforcement is going to get called up, you know, on all these fraud events or, you know, transmission shutdown events. [9:46] And their ability to see what's going on is going to be nil, unless the monitoring and safeguards are built into the AI. [9:56] And so that is a super urgent thing that it won't slow us down that much. [10:05] You know, that's another debate of should we slow down or not. [10:10] But deciding to put the safeguards of monitoring in, that's a choice we have. [10:15] And the only question is, do we do that before or after there's some large-scale cyber attacks or bio-attacks? [10:24] You take me to my next question, which is the debate about what to do. [10:29] There's a debate over, should there be a kill switch that is imposed? [10:36] Others are suggesting a total pause, as you note, on AI development. [10:40] And there are some who say companies can regulate themselves. [10:43] How do you believe AI should be regulated? [10:46] Well, unfortunately, most dangerous technologies, the government understands because they funded the research [10:55] or they're an important customer for nuclear bombs or rockets or things like that. [11:02] AI has emerged from the private sector. [11:05] And so the depth of understanding of these dangerous thresholds that were crossed literally this year, [11:12] bio-attack capability, cyber-attack capability, and others, you have so little expertise in the government. [11:22] And, you know, because it's cast in this nationalistic, you know, us-versus-them type thing, [11:30] you have people who are saying, well, let's just go full speed ahead. [11:34] It's completely irresponsible not to require every AI to have these safeguards and monitoring capabilities. [11:44] And it only is a question of do you do it now? [11:54] This may, you know, the effort to do that won't slow things down in a meaningful way. [12:00] After all, we're on this exponential increase. [12:03] So, you know, it ought to be common sense to build the safeguards in. [12:07] But, you know, we don't yet have people who look to the U.S. for leadership on tough issues. [12:13] Usually that's benign leadership bringing people in. [12:17] And they're a little confused when that, you know, that's not taking place. [12:22] Are you saying that the safeguards should be implemented? [12:25] Absolutely. [12:26] It's the world's most obvious thing. [12:29] I mean, you know, Anthropic reported on people trying to come in and create bioweapons or targeting capabilities using their AI. [12:37] There's all sorts of very sophisticated fraud things that you don't need, you know, people in Cambodia typing messages. [12:45] The AI is a thousand times better than that. [12:48] And so these negatives are imminent because they have really strong business models behind them. [12:56] And I want to be really clear, make sure I understand what you mean by safeguards. [12:59] You talk about putting monitors in place. [13:03] There's also a discussion about a kill switch. [13:05] Is a kill switch possible? [13:07] Is that realistic at this point? [13:09] The kill switch is, that's kind of a weird thing. [13:12] It's not enough to have a kill switch. [13:18] We're not yet at the point where they autonomously grab computers and, you know, can't be shut down. [13:27] So that discussion just sort of shows how non-technical various people are. [13:35] You know, I would never want to say I'm against a kill switch, but you don't, when you're trying to moderate the bad behavior, you need insight and records of what's being done. [13:50] So the kill switches alone would not be, not prevent these tragedies. [13:56] What would prevent tragedies from happening due to AI, specifically? [14:01] Well, I wish it was easier to distinguish good intent, like I want to design a new drug for a disease, versus negative intent, which is I want to design a tool of bioterrorism. [14:17] Because those are so close, we do need to monitor any sophisticated model and record exactly what's being done. [14:31] The United States ought to make sure this is done. [14:34] China will want to do this. [14:37] Usually, people expect the U.S. to be very concrete about what this looks like. [14:44] And you'll hear from the industry, it's a very doable thing. [14:48] It would not slow them down dramatically. [14:51] And so it's this, the safeguards and monitoring are separate from, would we be better off slowing down? [15:00] Anthropics CEO Dario Amadei recently called on AI companies to slow model development to try to mitigate risks. [15:08] Do you think self-regulation is sufficient or does it need to be accompanied by legislation in Washington? [15:14] No one thinks self-regulation is enough. [15:16] So there needs to be legislation passed in Washington? [15:19] Absolutely. [15:20] Absolutely. [15:20] You need law enforcement and the politicians to get into the discussion about what safeguards [15:30] and monitoring look like. [15:32] And that has to be a required thing. [15:36] And it will be a little bit of overhead for the industry, but not a dramatic slowing of what they're doing. [15:42] How much time do you think Washington has to take action before AI is ultimately out of human's control? [15:54] Well, a complete loss of control is very different than saying a bioweapon kills 20 million people [16:02] or that, you know, bank accounts get, you know, we care about things way short of the ultimate doomsday scenario. [16:16] And I'm afraid that's kind of distracting people. [16:20] It's, there is not a current risk of these AIs getting completely out of control. [16:26] The same monitoring that we put in to avoid bad people using AI is exactly the monitoring you want to see if the AI itself is initiating harmful activity. [16:39] But the, the thing that's urgent has to do with bad people using AI, not the AI going off on its own. [16:48] It has, you know, in a sense, I'd say it's good that these loss of control incidents have gotten people focused on AI, even though that's not the imminent risk. [16:58] If you were talking to lawmakers in Washington and you were to give them a timeframe, what would that be for them to come up with real concrete legislation to address this, to put the type of monitors in place that you're talking about? [17:14] Um, this, because of the limited expertise in the government, there'll have to be a construct where you draw on responsible members of the industry and have debate that, of course, the government has the ultimate power here. [17:32] And right now we don't have a way of drawing that together. [17:36] You know, certainly all countries should want these safeguards in place. [17:42] So anyone that says, well, uh, you know, China won't put these safeguards in place. [17:47] I totally disagree with that. [17:48] You know, you really, you have to show that you're willing to do this. [17:53] Then you can have a real discussion. [17:55] Just having two people say, hey, we have voluntary review. [17:59] What do you have? [18:00] That's doesn't do anything. [18:02] So you're saying there needs to be a global response, global guardrails put in place in order for this to be effective. [18:09] That's right. [18:10] It's common sense that you need to be able to see what these models are doing. [18:19] Their ability to attack networks is profound. [18:25] And we crossed that threshold this year. [18:28] And, you know, the reason I'm speaking out about this and, you know, instead of just being about, you know, eradicating polio or malaria, [18:39] is because I was stunned that outside the industry, as we crossed these thresholds, it was almost like people didn't understand it had happened. [18:51] Because the very straightforward safeguards, you know, we have voluntary review. [18:59] And that, that is, it's actually pretty scary. [19:03] When the people who are watching this, just to make sure I'm clear and they're clear, when you say cross these thresholds, you mean the fact that some AI systems have actually hacked into other AI systems. [19:12] No, the, well, the threshold of most concern is, can an AI design a bioterrorism weapon to kill millions of people this year? [19:22] Can an AI find flaws in code to attack and shut down electricity or hospitals or to scramble banking systems this year? [19:35] And so those capabilities, if you're not avoiding somebody bad, using these AIs to do those things, that's very, very risky. [19:48] On top of that, you have the idea that in certain of these scenarios, even the creators didn't expect the AI to be so autonomous. [19:58] And that's the longer term warning about, not only do we have to stop bad people from doing things, but we have to stop the AI itself from getting out of control. [20:12] Very important to work on that, but that's not the danger where you could wake up any day and see a massive attack. [20:22] That would be a human directed massive attack today. [20:25] You recently said you want to try to talk to President Trump about AI. [20:30] What will your message be to him about AI? [20:33] You know, I hope that given my life's work in the field, my saying how unique and different this is and how concerned I am will add to what he's hearing from other people. [20:49] And he'll say, okay, one of the great things I can do is convene people to talk about what these safeguards look like. [20:59] And, you know, I won't be saying we should have an overall pause or slowdown. [21:06] I'll be saying we should impose a requirement on all the AIs in the U.S. [21:13] And I expect, you know, China doesn't want bioterrorism. [21:16] They don't want big attacks by non-government actors. [21:21] You know, so I don't think this will affect the nation state competition. [21:29] As you know, President Trump recently posted, AI taking over the world, destroying humanity and all other things bad is a hoax. [21:37] Do you believe these warnings about AI posing potentially existential risk to humanity is a hoax? [21:44] No, it's not a hoax at all, but just because his initial reaction was that, he'll listen. [21:53] You know, he wants to protect the U.S. people. [21:56] You know, most people don't understand what's going on with AI. [22:02] They're not using the very latest models. [22:05] So what he said is fine, but I think after he hears from me and other people, [22:13] including that this does not handicap us in whatever he thinks the nation state race is, [22:22] that, you know, we will come to a consensus. [22:25] And as I say, some awful events would increase the urgency, [22:32] but getting the discussion in parallel about safeguards will help us get them in sooner than later. [22:40] We've been talking a lot about China. [22:42] President Trump's going to meet with Chinese President Xi Jinping. [22:46] What do you think is the best possible deal that the U.S. and China can reach as it relates to AI? [22:55] Well, there's nothing on the table today. [22:57] So they can, you know, I hope the tea is good. [23:00] I hope they smile at each other. [23:03] There's no proposal today to take this imminent danger and to learn from each other and work together to monitor and safeguard against these dangers. [23:14] So, you know, it's great that they're willing to talk with each other, but there's no substance to what can be done to have AI curb from negative activity. [23:31] You know, it's interesting because during the Cold War, the U.S. and the then USSR were able to cooperate on limitations around nuclear weapons development. [23:43] Is that a model that can be used right now between the United States and China to try to reach an agreement on AI? [23:51] Yeah, absolutely. This is more difficult than that was. [23:56] More difficult than nuclear weapons? [23:58] Yes. Nuclear weapons, access to fissile materials, the ability to see tasks. [24:04] There were things that were unique about it. [24:06] You know, even there, we could be doing better in terms of drawing China into some of those frameworks. [24:12] We've let a lot of that stuff expire. [24:16] But since World War II, the U.S. should be very proud that when there have been global challenges, [24:22] we, in a way that was benign, you know, where we had friends and allies coming along with us, you know, aviation safety, ozone layer, nuclear weapons, [24:35] a lot of things where we were played the fair actor, not aggrandizing whatever our unique strengths were. [24:43] And AI is one we're going to have to go back to this. [24:46] Hey, we're all in humanity is all in this together. [24:49] Big picture questions on AI. I do want to give you the opportunity to talk about the Jeffrey Epstein matter. [24:57] Obviously, you testified to Congress about that relationship in June. [25:03] And you said you had no knowledge or participation in Epstein's criminal conduct. [25:08] You've also said you regretted every minute you spent with Jeffrey Epstein. [25:13] What exactly did you regret? [25:15] Well, I was told by Jeffrey after I'd met him that he could help raise billions of dollars for global health. [25:26] You know, he wouldn't be paid. He would have no role. [25:30] You know, that proved to be a complete dead end. [25:33] So, you know, my spending time with him, like many other people, kind of helped his reputation. [25:41] You know, that's very regrettable. [25:43] You know, I really appreciated the chance to go in front of the Congress [25:48] and make clear, you know, go through every meeting I have with them, [25:52] which is about, you know, ten meetings over a couple-year period. [25:57] You know, I never, there were never any, you know, women in those meetings. [26:01] There's nothing financially. [26:04] Never went to the island or anything like that. [26:08] So, you know, if people are curious about that, I encourage them to read. [26:12] You know, I was there to answer any question that people had. [26:17] And you did acknowledge that you knew Jeffrey Epstein was a convicted sex offender [26:22] when you, for part of the time that you worked with him. [26:24] The Justice Department says there were more than a thousand survivors of his crimes. [26:30] What would you say to the survivors of Jeffrey Epstein's crimes? [26:35] Well, in general, we expect the government is involved in law enforcement. [26:40] And the fact that in the Florida case, the government failed to have a penalty [26:50] and they failed to see things that were ongoing. [26:53] You know, that's an incredible failure of the justice system. [26:56] A couple of other big picture questions on AI, if I could, and we're almost out of time. [27:03] I want to talk about the future and the younger generations. [27:08] How do you think AI should be used in education? [27:13] Is there any age that is too young to introduce AI? [27:17] Well, because you could use AI as kind of a friend or companion [27:24] in a way that would be always agreeing with you, [27:29] there is a huge risk that you're learning how to deal with people who aren't always sycophantic towards you, [27:39] that that could retard social development. [27:42] And so China has banned these kinds of uses. [27:48] And so we really have to distinguish between helping you understand math and personalized learning, [27:55] where it can be motivational and work with you at the right level, [27:59] and an emotional type relationship that, you know, probably does not serve your long-term interest. [28:08] How do you use AI in your everyday life? [28:13] Well, I'm very lucky that anything I'm curious about, getting very complex analysis and rich information, [28:21] the AI does such a fantastic job. [28:25] And, you know, it's helping me in the foundation work. [28:29] You know, what's the latest on malaria? [28:32] What's the latest on malnutrition? [28:34] You know, it's allowing the Gates Foundation, which will spend all its money over the next 19 years, [28:43] you know, we've got a bit over $200 billion. [28:46] That money will be far, far better spent because we're using AI to guide our work. [28:52] Are there any ways in which you professionally or personally avoid AI and say, [28:57] I'm not going to, you know, consult with it in this instance? [29:01] Well, I'm not using it for my personal relationships. [29:05] I'm not like, hey, do you like me? Hey, I'm a nice guy. [29:09] So I just never went down that path because I'm lucky in terms of my family, friends and everything. [29:15] I don't socialize with the AI. [29:19] It knows I'm a very serious person. [29:22] Well, let's talk about your family. [29:23] You have three kids. [29:24] You have two grandkids. [29:26] When you think about these challenges that the country, that the global community is facing, [29:31] as it relates to AI, what do you hope for their future? [29:34] Well, AI will have two phases. [29:38] It will have the phase we're entering into now where it has the potential to solve problems, [29:45] but it hasn't solved those problems. [29:47] You know, you're still, rent is too high, health is too expensive, [29:51] and yet it will be changing the job market and potentially empowering criminals [29:57] if we don't have the right safeguards. [30:00] So you have a period that's probably 20 years long where you have this adjustment. [30:10] Once you get through that, then a combination of humanoid robots and AIs [30:18] will provide, you know, building houses, growing food, a lot of things. [30:23] You know, you could call it an air of abundance. [30:25] There'll be a challenge there of meaning and how you spend time, [30:31] but you won't have shortages that we will today. [30:38] You know, that's for the younger generation to figure out what that looks like. [30:41] My focus is can we get through this period of turmoil without too much damage? [30:47] You know, maintaining equity. [30:49] You know, how can we get the person who would never meet a doctor in Africa [30:54] to have, you know, as good a health advice as a rich person or the farmer in Africa? [31:01] And so bringing this equity lens to us, you know, there are clear projects, [31:05] you know, they're already started. [31:07] That's very exciting. [31:09] So, you know, telling kids use AI in the right way, you know, [31:15] pursue what you're curious about, that advice is still timely, [31:20] but because of the drastic changes coming, my ability to predict what the world looks like, [31:27] even in ten years, is much less now than during my entire life. [31:32] I mean, this evolutionary event of a foreign intelligence that didn't come on spaceships, [31:39] it, you know, came in these computers, it's here, and it's improving at a mind-blowing rate. [31:47] Bill Gates, thank you so very much for being here. We really appreciate it. [31:53] We thank you for watching. And remember, stay updated on breaking news and top stories [32:00] on the NBC News app or watch live on our YouTube channel.