World Changing Technology In 2026 Fsi Stanford

TITLE: World Changing Technology in 2026 CHANNEL: FSIStanford DATE: 2026-05-06 ---TRANSCRIPT--- Welcome to our second in what is a new quarterly series that we’re doing here at FSI. I’ve been calling it, at least informally, the State of the World series. So the last session we did looked at the rapidly evolving geopolitical environment in 2026, where we seem to have about 10 years’ worth of news in the first couple months. It hasn’t slowed down since. Today we’re going to look at world-changing technology. And for those of you who don’t know me, by the way, I’m Colin Kahl. I’m the relatively new director here at FSI, although these types of jobs are measured more in dog years than human years. So it’s only been a few months, but it feels like a lot longer than that. This is a topic that’s near and dear to my heart. I oftentimes, among the FSI crowd, talk about that we’re not only living in a world of new geopolitics, but we’re living in a world of new technopolitics. That’s a phrase actually that I appropriated from Ian Bremmer, who’s the head of the Eurasia Group. But when I use the term technopolitics, basically what I mean is two things. One, that we are living in an era of world-changing technologies that will have profound impact on our societies, our national security, our prosperity, our way of life. And so we need to understand that impact. Here at FSI, we have some centers that are dedicated to doing exactly that. But technopolitics is also in recognition that these world-changing technologies are really originating not from the public sector, but from the private sector, and that these private corporations wield enormous international political power. They are multinational in scope. If their market cap were GDP, they would rank as G20 countries. Countries when you’re talking about companies that are worth trillions of dollars. And they have near-sovereign control over many of the digital domains through which we experience our lives. So they’re really important political and geopolitical actors in their own right. So when I say technopolitics, that’s what I mean. So we’re going to talk about world-changing technologies today, and I could not think of a better panel to help us think through some of these issues. The goal of this series is really to feature the world-class scholars that we have here at FSI, and this panel is a perfect example of that. So we’ll just go kind of down the here. So Drew Endy, that guy there, that guy. Drew Endy is a professor here at Stanford who studies synthetic biology. He’s an associate professor in bioengineering. He’s also a senior fellow by courtesy at both FSI and the Hoover Institution. He served on the US National Science Advisory Board for Biosecurity, among other national committees and task forces that advise the US government. And Esquire magazine recognized Drew as one of the 75 most influential people of the 21st century. You went there. Yeah. Yeah. Why not top 10, Drew? Anyway, OK. Andy Grotto, next to Drew, this guy. Andy Grotto is a research scholar here at FSI’s Center for International Security and Cooperation, CISAC, where he’s the director of the Program on Geopolitics, Technology, and Governance. His research focuses on national security and America’s global leadership in information technology innovation. Previously, Andy served as the senior director for cybersecurity policy at the White House, having the honor of working for both President Obama and President Trump in his first term. Not a lot of people fit that category. Andy survived in both places. Jeff Hancock, this guy, is a professor in our Department of Communications. He’s a senior fellow here at FSI as well, where he is the director of our Tech Impact and Policy Center, formerly the Cyber Policy Center, rebranded, remissioned as TIPP, the Technology Impact and Policy Center. He is a leading expert in behavioral sciences and the psychology of online interaction and the psychological aspects of social media and AI technology. And then last but not least is Jennifer Pan, who’s also a professor of communication and a senior fellow here at FSI, where she works most closely with SCII, which is our Center on China’s Economy and Institutions. Her research focuses on political communication, digital media, and authoritarian politics. She’s also a professor by courtesy of political science and sociology. So this is an august panel. So I think we should just jump right in. I’ll ask kind of a handful of questions to each. I will probably have follow-ups that try to bring them into dialogue. And then we will have plenty of time at the end for you to ask questions of the panel as well. All right, Drew, let’s start with you. You and I have interacted a bunch. I have learned so much about biotechnology from you, and I’ve heard you talk repeatedly about, you know, that maybe 2 years from now, 5 years from now, 10 years from now, we’ll be talking about biotechnology as an emerging general-purpose technology, much like we talk about AI as a general-purpose technology. And I wonder if you can kind of just flesh out that idea for this group. Yeah. So biotech’s about partnering with biology to solve problems. Biotechnology is ancient and modern. We’ve domesticated crops and animals over thousands of years. Stanford helped invent genetic engineering half a century ago. On the frontiers of a third wave of biotechnology is synthetic biology. The word synthesis means composition or putting things together. And so what happens when you not just edit natural living systems, but you compose living systems. So that’s just the three waves of biotechnology that are all intermingled and amplifying one another. Biotech’s gone to market, so it’s about $4 trillion of revenue globally. The US has about $1 trillion of genetically engineered domestic product. That’s about 5% of our economy. Some say— McKinsey says that goes to $30 trillion over 25 years. That would mean the majority of the physical inputs to our economy are biomanufactured. 1,000 years ago, our economy was mostly biomanufactured. The last 100 years, we’ve had a shunt through modern synthesis and chemical refining. And so the question is, how much do we go back, back to the future, but with a new, new biomanufacturing stack? Pulling that off is not trivial, but it’s in front of us. And so If we get to that, then you can begin to say, well, the majority of the physical inputs to our economy are biomade. Maybe we’ve got a general purpose technology, not for computing, but for manufacturing, right? What biology does is it takes energy, whether it’s sunlight or sugar, and controlled by information through the DNA sequences, it organizes matter, right? So you have a one-dimensional polymer DNA, and then that encodes the instructions for proteins that fold up into three-dimensional shapes and control the behavior of living systems. In the middle of this is a brand new technology called synthetic DNA chemistry, and that lets us take information in a database or however you want to keep track of it and put it into a DNA printer that makes DNA from scratch wherever you want to build the DNA. It’s— I’m hopelessly biased, but it’s impossible for me to overstate the significance of being able to manufacture DNA from scratch. It’s like how important silicon wafer manufacturing is for computing. But this is the stuff encoding all of life. So we know biology makes food, it makes fuels, it makes materials like cotton for clothing, it makes medicines. But that’s not enough to make it a GPT. To make it a GPT, we’d have to be able to grow other stuff. And some quick examples, and then I’ll just Paused a brewing process you could use to make beer or wine or medicine over since 2011 is also now capable of being reprogrammed to make explosives. Right. So suddenly you’re thinking about a biomanufacturing process for hard power. There is AI-assisted protein engineering that’s now allowing for enzymes to be made from scratch that are catalyzing the deposition of metal oxide semiconductors. So what if there’s an alternative path to making computers? It’s not top-down lithography with expensive factories in Arizona and machines from the Netherlands. It’s bottom-up self-assembly. The biology is not the computer, it’s the manufacturing process that’s providing a more distributed, more affordable way to making things that are currently seemingly difficult to make. So that’s, that’s what’s in front of us. And, you know, I’ve been in the business for 20 years and it’s, It’s amazing how it’s playing out. There’s— it’s still mostly, you know, ahead of us, but it’s happening. So I think one of the most fascinating things about biotechnology and especially this age of synthetic biology is it seems like there is a high degree of interaction with another world-changing technology, which is artificial intelligence. So you have biological design tools, you have foundation models, which are also getting better at biology. It seems like we’re reaching this moment where, you know, we’re turning— increasingly turning living organisms into zeros and ones, and we’re turning zeros and ones into living organisms. And I wonder if you could talk a little bit about the interrelationship between AI and the biotech revolution that you just talked about. I think you’ve set it up exactly right. And I want to add energy to the equation too. But let’s start with compute and AI. If you had a DNA printer that could build any DNA you want, you quickly confront a puzzle. What do I say? You know, like we can learn English or Chinese or French or pick your favorite human language, but speaking in DNA is tricky. Like I know a few sequences, TAATA CGACTC ACTATA GGGAGA. That’s a little sequence that causes DNA to be read out. It’s called a promoter. And I made that my computer login password. By the way, he just makes up a different sequence. No, no, no, no, no. How many times— I’ve heard him say that. It’s never the same sequence. T-A-A-T-A-C-G-A-C-T-C-A-C-T-A-T. M-I-S-S-I-S-S-I-P-P-I. That’s as much as I got. But like, I only know a few because I really made them my Gmail account or my computer login. And so imagine you have a machine that can print DNA from scratch. You’ve got to figure out how to feed valuable sequences into that machine. So here’s where AI in one example is extraordinary. This is the work of Brian He, a junior colleague here in chemical engineering and data science, you can take natural DNA sequences found on Earth, arrived at by evolution, and you can train a large language model on those natural sequences. This is what Brian and colleagues did. It’s a tool called EVO and now EVO-2. It’s a large language model for DNA. If you can figure out how to prompt it, it emits DNA sequences. Now, the puzzle with these DNA sequences is we don’t know how to read them very well either. And so we have to build them and test them to see what they do. But, but it’s so exciting. We’ve never had a tool that scales for generating higher quality information going into the DNA printers. And we can use this over and over again to make better enzymes, better therapies, better diagnostics. Of course, you could flip it around. You could make worse toxins, worse pathogens. So we’re going to have to mind the risks by a lot, you know, but that’s, that’s just one quick example. The other examples, of course, are using AI tools to model biology, right? So over the last half century, not only have we accumulated a lot of DNA sequence, we’ve accumulated a lot of three-dimensional structures for proteins, like the public investment in solving protein structures over the last few decades, probably about $100 billion worth of experiments. Google DeepMind and AlphaFold took that database and built AI tools for predicting the shapes of all proteins. The thing in front of us that’s going to keep us from fully unlocking the power of AI to help with biocontent is we’re going to need more data. We have really good data for protein shape and for DNA sequence, but for any other state variable describing how living systems are operating, we haven’t collected it on a systematic basis. So one of our policy recommendations is if we want the United States or NATO or you name it to be world leading in AI-driven biotechnology, we have to have large language laboratories. For generating data that feed into the large language models. So that’s thing one. Thing two is energy. As a civilization, we need primary energy, electricity, and then we need embodied chemical energy in the stuff all around us. And we need more of it. Maybe it’s 20 terawatts of primary energy today for civilization and 16 terawatts of embodied chemical energy. Most people will think about energy skip on the embodied stuff. They just think about the liquid fuel and the gas and the electricity. But both are important. If we get to a 2050 and the world is still inequitable, like the energy gradients that exist today, we’re going to need to go from 16 to 24 terawatts, 8 more terawatts of embodied chemical energy. Where’s that going to come from? More agriculture? More shunting of petroleum into stuff? Or how’s that going to work? And of course, we need more primary energy. You know, think about the methane that gets burned to make fertilizer to run agriculture and how that’s going to be an increasing puzzle with the geopolitics of supply. So there’s a massive intersection that I think we’re sleeping on a little bit between biology, biomanufacturing, and energy. The most exciting opportunity in this area is what I call electro-biosynthesis. That’s a mouthful, so just think of it as e-bio for short. E-bio, like an e-bike. You’ve got an electrical bike, right? So, so e-bio is you take electricity however you can generate it. My favorite is peak solar when you don’t know what to do with the excess solar electricity and you run that through a process that pulls carbon out of the air and turns that into an organic molecule like formate, which you can feed to microorganisms. Instead of glucose. So you go from photovoltaic-generated electricity, atmospheric carbon, and then a type of molecular battery, formate, a chemical that’s storing that energy you can put into biomass. It’s the physical analog of agriculture, natural photosynthesis, but instead of a leaf, you’re going through a photovoltaic panel into a biomanufacturing process. This has not been developed at scale. The leading companies are around Pacific Northwest National Labs, and ADM in the Midwest is starting to pick it up. But I think of this as important as learning to burn methane to fix nitrogen to get synthetic fertilizer over a century ago. It’s a massive opportunity to shunt primary energy production into embodied chemical energy. So it’s both the bits and the joules that are converging with biolab. Well, the other place where bits and joules intersect, of course, is, is AI, where data centers are extraordinarily energy hungry. Jen and I actually were at an event earlier today with the House Foreign Affairs Committee staff that Hoover hosted, where, you know, the discussion was how do we win the AI race, right? Which everybody’s obsessed in DC and Silicon Valley about winning the AI race. Annie, I want to turn to you. You know, much has been made about the race between the United States and China on AI. I wonder how you think about the race and why people are so obsessed or should be obsessed or maybe not obsessed with who wins. Race is not the best metaphor, but it has its virtue. It speaks to what kind of preparation is needed in order to perform well. But sticking with the sports metaphor, I actually prefer to use a multi-sport event like a like a decathlon, where athletes compete in numerous individual events, each drawing on a slightly different but often complementary set of skills to win. AI competition is a multidimensional contest where ultimate success depends on a high performance in multiple different events, each with its own preconditions for success. And understanding the competitive stakes for each of these individual events and how they contribute to overall leadership, I think, is essential preparation for designing policies that put the United States in particular in the best possible position to succeed. So my running count of events is 7, a heptathlon. There’s an event at the frontier, whose computing power and algorithms are most advanced with the ultimate prize, according to some, Be an achievement of artificial general intelligence. Whoever wields AGI, to paraphrase Dario Amadei, the co-founder of Anthropic, will have access to a country of geniuses in a data center. A staggering and probably world-changing source of national power. Another event, applications. Who does better at using algorithms, whether frontier or not? To power the killer apps that consumers and businesses ultimately embrace. Customers aren’t usually buying a model. They’re buying the app that uses one or more models to build a solution that helps the customer solve or achieve some task. Producing killer apps is not the same as using them. And so those roles— vendor and customer that contribute to national power in different ways. There’s widespread use of AI diffusion that could lead to productivity gains, whereas producing the killer apps for AI, you know, enriches investors, powers research and development innovation cycles, and gives government some leverage over who else can use the app, for example, via export controls. So diffusion is another— is the third event. Who reaps the greatest productivity gains from AI use diffusing across the domestic economy? While at the same time, and I think this is really important, while at the same time managing the social disruptions caused by shifts in the labor market, right? So diffusion of AI across the economy that results in massive social disruption that causes a backlash, that’s probably not a recipe for victory. Another contest, military adoption of AI, harnessing AI for war. Warfighting. It’s a crucial but distinct competitive event as well. AI could flourish in the civilian sector but flounder in the military, or at least diffuse at a slower pace. The reverse, of course, is also true. The military could initially outpace the civilian economy in adoption. This has happened before in areas like semiconductors and GPS technologies. And of course, both sectors, civilian and military, could fail to adopt a new technology similar to the failure of the US shipbuilding industry at the turn of the 19th century to transition from building sailing ships out of wood to steamships from metal. Manufacturing, another event. Some AI innovations will manifest as apps where the marginal cost of producing another unit of software is effectively zero. But others will take the form of tangible products that have to be manufactured at at some cost. You know, think new medicines, new materials, new product designs, and so on. And a country that can’t produce these things for themselves is missing out on a potential source of valuable economic and social benefits from domestic manufacturing. It’s also more geopolitically exposed to economic coercion than a comparatively more self-sufficient rival. Market share is the sixth contest. Has to do with whose AI stack, computing power, models, applications, and so on captures the most global market share. This matters because revenue is good for national growth and helps fuel R&D cycles. It also matters because technologies are often embedded with the values of their originator, and global diffusion of AI with American characteristics, or at least liberal democratic characteristics, as opposed to CCP characteristics, is probably better at sustaining and maybe even growing US influence globally. And of course, global market share could also make customers dependent on the provider, potentially conferring the provider with a source of coercive influence over the customers. And that possibility is what you see animating much of the growing concern globally around AI sovereignty. On the last contest, this has to do with governance, whose values and interests are best reflected in the ways that people and governments around the world use and govern AI. Each of these contests, frontier, applications, diffusion, warfare, manufacturing, market share, and governance, they all contribute to leadership in AI and success. And one of them can often contribute to success in another. But success in any one of them does not necessarily guarantee success and success in the other domains. Each of them makes an appearance in US and Chinese policy, I would argue, but I think only China, so far at least, has pursued an all-of-the-above strategy. So this is really— I think this notion of multiple races, which is something you and I have talked about, multiple races, multiple events, however you want to characterize it, is interesting because, you know, my sense is US companies remain 6 to 9 months ahead at the frontier in terms of models, conservatively more and better compute resources. So if you measure it just on that race, US, the United States is doing pretty well, even though China is a fast follower. But China is competitive everywhere else, right? They’re, they’re super competitive on the apps. They’re super competitive on the diffusion of good enough AI and related infrastructure. They are arguably ahead of us on the industrial applications of AI. And the jury’s out on AI governance because The Trump administration kind of backed away from that, although interestingly, I think it’s going to be back on the agenda when Xi and, and Trump meet in Beijing shortly. But I want to lock in on the one piece, which is diffusion, right? Diffusion, which is whose, whose stack, as you call it, right? Whose models riding on what chips in what data centers with what surrounding infrastructure basically determine how people around the world access AI. Can you talk a little bit— I know a lot of the work that’s been done at GDG has focused on this diffusion question. You, you talk a little bit about what that research tells us about where the United States and China may have the edge or, or, or not. Like, what determines that competition? Yeah, I think, you know, the foundational challenge for the United States and its, and its companies is that, you know, China is not only a formidable competitor on on price, but increasingly quality. And that means that touting sort of platinum-plated AI, if it’s higher cost, that platinum plating may not be good enough of an improvement over China’s gold-plated AI to rationalize that higher price that will often be in place. And so the challenge for the United States is to think about how to promote American technology on quality and price. It’s got to be cost competitive as well as continue to compete on quality, but expanding the idea of quality to include geopolitically, right, who would you rather align with? And we’re at a difficult moment, I think, in American foreign policy where That choice, I think, is difficult for many countries to consider making today. And our research is showing that the countries, they want sovereignty. Now we can unpack what that means. It means different things in different countries. But usually it means having some domestic control over what rules apply to technology, preferably their own legal traditions, what cultural traditions are brought to bear to inform those rules. Again, that’s a reflection of the domestic governance ecosystem. Countries care a lot about the apps being able to service the local market, which may mean having apps that are built in local languages, that solve problems that may be unique to that a particular marketplace. And I think— so those are just some of the preliminary findings that we’ve uncovered. I wouldn’t describe any of it as particularly earth-shattering, but I think we see it documented in country after country and after country. And what I think we need to do better at here in the United States is develop a better sense of strategic empathy for what this debate looks like from the standpoint of middle powers and emerging economies around the world. Yeah, I mean, I think one of the interesting things from the research that, that you all have done— and full disclosure, I’m the faculty director of the program too, so I’ve been involved as well— is that the United States and China basically pursue different models, right? The United States is trying to unleash the market, flood the zone, but the market will serve countries that have large populations, large consumer bases, or have a lot of capital, a lot of energy, something else. So like Google, Microsoft, AWS, the Frontier Labs, Anthropic, OpenAI, they will invest in these places. But China will invest in a huge swath of the Global South that doesn’t actually meet that criteria, and they’ll do it for strategic reasons. And so I think one of the questions around diffusion is how serious the United States will be in basically addressing this gap where the market is not serving. You know, a lot of the conversation about AI safety focuses on risks like cybersecurity risks. We’ve seen that with the, you know, the holdback of Mythos Preview, which is this highly capable cyber model that Anthropic has developed. There’s also concerns in the bio front about whether AI could help, you know, uplift malicious actors who want to create pathogens, you know, create some more dangerous version of COVID or worse. There’s all sorts of science fiction concerns about rogue AI. But there’s also just— if we, you know, I have kids. And I worry about how they consume digital information in all forms. And so I want to talk, Jeff, to you about that because so much of your research is about the impact of digital platforms and technologies on persuasion. And one way in which AI could warp, especially politics and democracies, right, is through propaganda, through censorship, through shaping, you know, the information, misinformation, disinformation that shapes our overall perceptions of the world. So I guess the question for you, Jeff, is how concerned should we be that AI laced through digital environments will shape our perceptions of reality, politics, and just our way of life? Yeah, answer: very concerned. End of story. So actually, I think there’s a lot of different things to unpack there. So one is we now look back on social media over the last let’s say 16 years, 2010 is when it becomes very mature, and there was a lot of optimism about what social media was gonna do, was connect us and allow people to grow and develop in new ways. After 2016, there was a real turn towards we’re worried about a lot around mental health. One of the things that kind of went missing was this idea that when you consume a lot of media, it can shape your understanding of the world, and in some ways that is what persuasion is. And so I’m in the communication department, as is Jen. Jen will talk a little bit more about propaganda. She knows a ton about that. But in many ways, communication is persuasion. And what we’ve come to realize is that a separate concern from mental health questions is shaping young people’s, especially, but also adults’ views of the world. So if you’re a young person, a young boy, and you start getting access and exposure to say the manosphere, which is a very nihilistic sort of space about, you know, women are not there for you, they’re to be, you know, to serve you, that you’re controlled by certain groups. It’s a really horrible space, but if you’re a young person sort of developing your identity and understanding, that’s gonna shape your understanding of the world and the way it works. If you’re a young woman that’s looking at things, you see people putting on makeup all the time and coming in very heavily dressed, You may come to understand that the world is one in which you should wear makeup all the time. Both of these aren’t necessarily going to affect your mental health, like anxiety or depression or loneliness, but they shape what you think the world is like when you go out into it. I think that is something that we’re coming to grips with now. And it’s not necessarily the algorithm, although I think it plays an important role. It’s the content that is served on those. And, um, Colin mentioned something in the beginning when he was talking about technopolitics, which is money. The amount of the economy that these platforms consume is part of what is driving these different kinds of content to be served to young people and adults alike. I just recently received a notice from Fidelity that the index fund in which I’ve invested is no longer legally a diversified fund because companies like Meta and Apple are too large, and so they make up more than 5% of my portfolio. So it’s no longer diversified. The reason I bring this up is that that content gets in front of us because we engage with it, and it’s not necessarily because we want to see it. It’s because in some ways our systems are being hacked. We look at negative content because potentially the way we evolved was to pay attention to that. Now we’re transitioning into an AI world in which there’s a lot of interaction there. We’ve been able to use AI to change the way people’s feeds work, where you get rid of some of that really horrible and polarizing content, and people like it more. They actually enjoy using those systems. They’re less polarized afterwards, but they use the system— in, in one case, Twitter— about 4 minutes less a day. And that doesn’t sound like a lot. That’s $1.4 billion for Twitter. So, imagine that Elon Musk is like, okay, I’m gonna do the right thing, Jeff, and we’re gonna like reduce all that bad content. He would likely be sued by stakeholders and shareholders. Like, you’re not allowed to do that because we will lose money. And if you think of that then at Meta scale, it’s even worse. And so there’s these incentive— economic incentives that are bumping up against our psychology that’s causing a lot of these problems. I mean, it’s really interesting because, I mean, look, a lot of us at FSI, we travel the world. But even those of us who travel the world, the time— I would say 90% to 99% of what we understand about the world is mediated through digital environments, basically what we’re seeing in our— on our phones or on our laptops or whatever. I might— My daughter would be horrified that I tell this story, but I’m going to tell it anyway. She’s a freshman in high school. 3 years ago, she came into the kitchen. She said, “Mom, have you ever thought about getting this dietary supplement? Like, you know, I’ve heard that it’s really good for you,” and I don’t even remember what it was. And my wife was like, you know, she asked her, she’s like, “Did you see a commercial on YouTube or something?” She’s like, “Mom, nobody watches commercials.” “No, I don’t— commercials. I’m not—” It was from an influencer. Yeah, TikTok, of course, a commercial in a different, in a different form. All right. So a lot of your research, you know, we are very concerned about how much time our kids spend on their screens, the impact that that has on their mental health, addiction, depression, anxiety, self-esteem, all these things. You’ve also done some interesting research, start or starting to look at research at various efforts to ban social media for certain, for teenagers. These social media apps obviously already are basically driven by AI algorithms that feed you content that keeps you engaged. What should we be learning about, about the early attempts to knock kids off social media platforms? And what does that mean for the kind of the AI age? Yeah, absolutely. So December 10th of last year, Australia passed something called the SMAA, the Social Media Minimum Age Act. which prevents kids under 16 from going on social media. How many people are aware of that ban? Wow, okay, really amazing. So, my group is the lead academic evaluator for that. We work with the Australian government to assess how is this ban affecting young people and their families. And I can tell you some of the early things that we’re seeing is compliance. It’s a weird law. Unlike here in the US where we have laws around accessing, for example, alcohol. If a young person goes and gets alcohol, that’s actually illegal. In this act, all of it is aimed at the platform. So if they go online, it is actually not illegal. What is illegal, violating the law, is that if a platform allows a young person on. Now I can tell you some of the unfortunate early news is that it does not look like the companies are complying with the law. They’re doing the bare minimum and they’re hiring other companies to try and you know, see if a young person’s on there or not. But the eSafety Commission, which is the part of the government in Australia that manages this, is really clear now that the platforms look to not be complying with the law, which is allowing a lot of young people on. There’s some good news, which is that about 35% of young people, and this is according to them and their parents, are not using social media. And maybe that sounds like a failure because the majority are still on, but actually this is the world’s It’s the first media policy like this ever. And we’re only about 4 months in, so to see that 35% are actually adhering to it is pretty amazing. So the early part of it is that there’s some positive signs there. Some of the negative is that in our early interviews with young people who are very different in their responses, 14 and 50-year-olds are angry. They’ve been deplatformed. It’s not through any fault of their own. They’re like, adults are the problem, and you’re kicking us off. So I think one thing the SMOD did that wasn’t thoughtful was kicking the 14- and 15-year-olds off, even though they were using it a lot. 10- to 12-year-olds, yeah, this sounds like a good idea. My parents think it’s important. My older brother who’s 17 or 18 thinks young kids shouldn’t be on it too. So I support this sort of thing. So that’s one thing. Another one that’s more worrisome is that when we ask them, OK, let’s say you do it here and you don’t spend as much time on social social media. What are you going to do? The goal is to go in the footy fields, as the Australians say. That’s not what they say. They’re going to go hang out more with a chatbot, go hang out with ChatGPT. So one potential for some kids is we’re moving them from a space where they’re looking at other and interacting with other humans to interacting with a chatbot. Lots of concerns there, obviously. If we go back to your daughter and persuasion there, she was being persuaded by a human and when we move them into these chat spaces, they’re being persuaded by bots. And I’m just gonna end here with two things, one like exciting and one quite worrisome. The exciting is that these LMs are very persuasive, probably more persuasive than talking to a person. And we just had a talk over at the center today where young people are going to advice for interactions they’re having that are quite difficult and they’re being convinced by LMs to make changes to the way they’re interacting that have been really helpful for them. Young people are trying to figure things out. We had a colleague, David Rand, just show that adults that admit to a bot that they have a conspiracy belief, say that 9/11 was an inside job— wait, it wasn’t? Yeah, I know. I’ll get you to talk to the bot later. But this is for the next panel. These LLMs can actually convince people to move away from their conspiracy theory. And the way they do it is not through rhetoric of emotion or things is that it’s persuasion through facts and evidence. It’s kind of amazing. So there’s actually ways in which this can be really powerful. But what if instead they’re talking to Grok 4, which has very unique ideas about maybe the way the world works, or they end up talking to Deepseek, which has very different values than, say, values of ones developed here? Who among us has not turned to Mecha Hitler? You don’t know what I mean. You’re going to have to take a look at that. You’ll have to take a look at that. I— well, this point of like, you know, serving up facts and evidence, like, whose facts, whose evidence? And Jen, I really want to turn to you on this because I think there’s no— any of us who’ve interacted with these LLMs on a regular basis, could be Gemini, could be ChatGPT, could be Grok, could be Claude. They are highly persuasive. And it’s easy to see how they can shape perceptions. And you’ve done a lot of work at the intersection of digital technologies and digital authoritarianism. And there’s a lot of concern that AI may enable digital authoritarianism, enable propaganda, censorship, surveillance. And so I wonder, in the China context, what you have seen, Jenna, on the relationship between digital technologies, how the Chinese Communist Party is thinking about AI, and the relationship between those things and digital authoritarianism. Authoritarianism. Yeah, so speaking specifically on China, I think the starting point is that the goal of the Chinese Communist Party is to survive in power. It’s always been its goal. It’s its goal today. It’s going to be its goal in the future. So from that premise, what, what is— how do we think about generative AI? And I think there’s been a lot of the conversation has focused on political control side, so surveillance, censorship, propaganda. But I think for the CCP, Chinese Communist Party, today, the core of generative AI is economic. It’s as economic stimulus. So since the Deepseek moment and beginning of 2025, you’ve seen this state push as well as bottom-up excitement around AI. And you see the diffusion and adoption across sectors. So I think that aspect of generative AI for authoritarian regimes shouldn’t be understated. And especially for the Chinese Communist Party, that AI can be an important stimulus for economic growth and economic development. That said, on the control side, in terms of, yeah, surveillance, censorship, propaganda, it’s not clear to me that it’s going to lead to an in-kind change to digital authoritarianism. Digital authoritarianism was in China before generative AI. Uh, computer vision and conventional machine learning were being used for mass surveillance. Censorship was being delegated to social media companies that were using all sorts of conventional machine learning to enable that. We— before GenAI, you see the Chinese Communist Party, uh, engaging in algorithmic competition for the public’s attention. So it’s not clear to me that GenAI is going to lead to some sort sort of phase or in-kind shift to digital authoritarianism in China. I think where it might lead to broader changes is in other contexts for less well-resourced political actors who previously may not have had the resources to engage in sophisticated information operations or to conduct surveillance at scale. That now potentially are enabled because of these technologies. So in other words, China was already there. Yeah, maybe at least in more efficiencies. But I think where at least the greater change is in other actors that previously didn’t have these resources. It’s really important because I think a lot of times, you know, people talk about China and the Chinese Communist Party somehow trying to export its ideology. And I think it’s much more accurate to say they are exporting their methodology. Their methodology of digital authoritarianism. Maybe agree or disagree. If, though, the CCP’s primary goal is political stability and control, there are ways in which frontier AI models dial up the capabilities that existed before the ChatGPT moment. So maybe it’s not a step change. It’s just dialing up the ability to do what they were already doing of surveilling their population, censoring what happens on the super apps, engaging in propaganda to shape perceptions, all of that. It just makes it more. But I wonder, is there also, though, a way in which it runs the risk of loosening control, especially because, you know, the Chinese models are not at the frontier, but they’re 6 months behind. And the key distinguishing factor is all the leading Chinese models are open weights, right? So you can download them and access them on your own infrastructure and you can find tune them and change the guardrails and do a lot of stuff with them, that can be extraordinarily empowering for citizens, right? And so at what point is the Chinese Communist Party going to crack down on powerful AI in the hands of their citizens out of concern that this will empower citizens to do things that undermine their control? I think the crackdown point will only happen if the risks to political stability are out— outweigh the economic benefits. And I really don’t see that at all at the moment. In 2023, China was one of the first countries to pass systematic AI regulations, so the 2023 Interim Measures on the Management of Generative Artificial Intelligence Services. And what that regulation did was basically extend China’s censorship regime from existing communication channels to generative AI. And in tests that we’ve done on China-origin models from 2023 to 2025, we see very strong compliance by Chinese AI companies. So when you prompt models on areas that are under the purview of the government regulations, these models will refuse to respond at much higher rates than US models. It’s— and this is not training data, not technical capabilities, because when we prompt them on areas outside of the regulatory control, there’s no gap between China provenance and non-China provenance models. So we do see this extension of China’s censorship regime to AI companies. Now, as you say, most of the China models are open weight, so anyone can download them, deploy them privately, potentially fine-tune them or ablate them to circumvent some of these censorship, censorship modes that have been implemented on them. So I think it’s definitely possible that people or civil society organizations will do this to try to circumvent censorship. Yeah, but I think the stronger imperative right now is still the— on the money side, but more from economic development perspective. These models are going to help me make money. Maybe they’ll also reveal information that we don’t know. But that imperative seems weaker at the moment than the economic one. It’s really interesting because, you know, when during the social media, the heyday, like the idealism around social media, right, there was a sense that an open internet and social media would essentially break down authoritarian regimes. I think what China proved capable of doing is erecting the Great Firewall of preventing the information and censoring the information coming in that is presenting the average person in China with a much different version of the internet than you or I would experience here. I wonder, though, would they worry that these two— I mean, we know the one place where the AI tools are getting the most advanced is in software engineering and coding. And one of the reasons Anthropic held back Mythos preview was because of its ability to identify vulnerabilities in basically every piece of software and operating system, even those that have been around for a long time. I just wonder whether the CCP is going to start to worry that people are going to start to hack through the Great Firewall and the other systems of surveillance? Or is that just something that authorities in Beijing you don’t think are currently worried about? I mean, I’m sure there is something that they’re considering. And I don’t think we can predict how policies will change. It depends on if there’s, for example, some sort of mobilization that’s enabled by AI that challenges legitimacy. But at present, and based on prior empirical work, what we see from the domestic Chinese population is that there’s limited demand for uncensored information. And so in a context where people don’t know what they don’t know, how are they going to seek that information? So even if you’re interacting with an ablated or fine-tuned or just ChatGPT or Claude, are you going to actually be asking the questions that are kind of censored politically? And I think more generally, across countries, political information is just not consumed that much. So I think to the prior discussion, yes, most of our experience these days is mediated digitally. But very little of that is political. And we’ve done research on this showing that for the Chinese Communist Party, propaganda has actually gotten more difficult with digital technologies, because now they’re competing with all these influencers for attention. They can censor very effectively, but can they reach the public broadly? And that’s become more challenging. So I think a priori, we can say very little about whether a new technology is going to be good or bad for authoritarianism. I mean, essentially, if you’ve depoliticized the population, they’re not going out to seek information that would be subversive. It’s really— it is interesting how the models cough up different pieces —information. Andy made reference to the models baking in certain values. We ran an exercise here at the Master’s of International Policy program in a class that I oversaw on foreign policy where I had them do a South China Sea simulation of a US-China crisis in the South China Sea. They wrote recommendations for the president, and then I had them feed their memos to one American model and one Chinese model to ask how China would respond. and the models gave very different answers on territorial claims in the South China Sea and who was the aggressor and who and how the Chinese government would likely respond. So that, that the value is being baked in is real. I do want to circle back on the, on one China question to, to get Drew to opine on something because Drew, we roped you into a Track 2 conversation with the Chinese on, on AI national security risk. One of the risks that people worry about the most is on the bio side. Right? I mentioned this kind of malicious uplift of, you know, the ability to design pathogens. A, how worried are you about that risk? And B, is this actually an area where the United States and China might be able to usefully collaborate? We know that it now looks like Xi Jinping and Trump may actually have AI safety on the agenda. I suspect that was kickstarted by the Mythos issue. But is bio an area too where the two superpowers might, might be be able to avoid a race to the bottom? Yeah, it was a great conversation in Munich in February with our colleagues from China. President Trump at the UN General Assembly last year said something, paraphrasing, call upon all nations to end the scourge of bioweapons. I’ll take that statement at face value and want to drive a truck through that. Yeah. We’re blessed to live in a world today where no nation-state brags about an offensive bioweapons capacity. I do not want to live in a future where nation-states are remilitarizing biology with the tools of the 21st century. It’s just, don’t put me down that path. It gets really bad really fast. I’m much more concerned about nation-state programs. So if I had a magic teleprompter and was programming what was coming out of the two presidents’ mouths, you know, in the near future, I’d want them to be shoulder to shoulder saying, Beijing and DC are cooperating on ending the scourge of bioweapons. That by itself, depending on who you talk to, isn’t worth very much, but I’ll take it because it gives me executive air cover to work with colleagues over the next 15, 20 years to really get biosecurity in place. If you wanted to wish for more, I’m happy to wish for more. I’d love to see Beijing and DC agree to a duty to notify in the event of an outbreak of a pandemic potential pathogen. That’d be better than nothing, right? And we could just start building on that. Now, back to your question, does AI contribute to biological risk? Sure. And you mentioned, you know, one of the three modes of risk. So one is you’re uplifting people who aren’t expert at practicing biotech. A lot of people who have looked to cause harm, they might think about biotechnology, then they see how difficult it is to do anything with biology and they’ll pick an automatic rifle instead, right? And there’s being rational actors. So if Claude or whatever makes that easier by upskilling or uplifting, you open the aperture, the landscape of unilateral actors who could misuse biology. Okay, not great, but okay. Second thing is you can use computational methods to make novel toxins and pathogens for which we don’t have detectors, medical countermeasures, or vaccines. You just work around stuff, and that would be bad too. So those are the two major threat paths. The other threat path is a little bit subtle, but returns to my biggest concern, which is nation-state BW. And so here’s the third path. Rhetoric around risk in biology leads to the possibility of nation-states trusting each other even less. And if you have a BW program, I’d better have one too. And this is the historical pattern that played out 100 years ago that led to the militarization of biology leading into World War II. Nothing good came of that. And so this, this heated rhetoric around AI risk that’s being lobbed around carries a type of communication risk that could be destabilizing at a nation-state level. And that’s why I’d like to put an end to that. That’s something we should be able to stop. But, but the AI community is so good at getting attention right now. Forgive me that, that, you know, you get attention for saving the world, promising to save the world, and we might destroy the world. It’s like Tony Hawk on the halfpipe of salvation and doom. And when you get all the attention, then you can get all the resources to do things. But when you get too much negative attention, you need to discharge that liability. And sometimes I feel like the AI liability is being discharged into Bioland. Large language land, la-la land, is taking the liability and putting it on me without really wanting to solve biosecurity. So you could take one further step. It’s like, well, why are we worried about toxins and pathogens? Why are we worried about novel toxins and pathogens? Why are we worried about more actors deploying novel toxins and pathogens? The answer is pretty simple. We haven’t secured biology yet. So it’s good to call attention to the risks. But if we don’t actually then move out and sustain efforts to be vigilant and secure biology, I don’t, I don’t really want to get pulled into those conversations because I think it’s, it’s, it’s dangerous in a different way. Yeah, I mean, it strikes me we also need to lean into the technology for, you know, biodetection for sure, for, you know, new therapeutics, new vaccines, new cures, but also just being able to identify, you know, patient zeros a lot earlier. And what’s really exciting is you can use biology and AI together to do that detection to create biological intelligence like we created geospatial intelligence. And let’s start with that. Yeah. All right. Enough from me. You get— you still get more of them. So we have some mics here. Please, if you want to ask the panelists a question, step up to the mic and ask an actual question. Make it relatively brief. And if it’s directed at a particular person, let them know. So over to the audience. Come on now. Otherwise, I’m going to have to get them to ask each other questions. Go ahead, sir. There’s a microphone right behind you. So Jeff, there’s a question for you. Okay. About social media. A lot of people believe that the advertising business model is really the fundamental problem with the way social media operate. And if you could come up with an alternative that would still give access, financial access to people so they wouldn’t have to pay a lot of money for it. And instead of surrendering their attention, they would pursue their interests, look for friends, look for compatible people. What’s your take on that? How big an issue is that? And is there any way to fix that now? Thank you so much for that. Colleague and I were just reviewing Facebook’s— Meta’s, sorry, revenue. And it’s over 94% of the revenue of revenue comes from advertising. So, it’s not like this is a diversified thing. All of it comes from advertising. I think we kind of like intuitively know that, but actually to understand what— there’s this trillion-dollar market that’s been captured by a small number. I’m not an economist, but as a psychologist, I can say that most people do not like the advertising model and at the same time, they don’t wanna pay for things. And so, when people, almost everybody will say, When it comes to social media, they’re worried about their privacy, and when it comes to AI, they’re worried about their privacy. What I think people are often saying is that it’s actually a fairness problem. What I mean there is that you’re building this model of me, all of my data, and then you’re essentially selling it via advertising. And what you’re getting out of that is not equivalent to what I’m getting. It’s not a fair trade. And so I think there’s a massive fairness question here. And I love your question because right now it’s getting asked anew in terms of the AI platforms. Right now they’re not using advertising. There was one try that failed pretty miserably. And the economic— I think the question for social media is sort of gone. It’s passed. There’s nothing we can do about it. And I think that they have decided that as well. And this is their model and they’re going with it. The, the AI platforms are kind of in that 2010 to 2016 era where it’s— there’s decisions that are going to be made here. And I really liked Andy’s point about values. Right now, the value system for AI in some sense is once you get past chat is can you help me with something in the world? So I really like using Claude Code this year. It was a sort of my ChatGPT moment of 2026. Where it’s like helping me do things. It’s not about me talking to Claude. Actually, that’s a very small part of it. It’s about going out and doing some work for me. So actually, the incentive there is to get me off the platform pretty quickly and have the platform go do work for me. If that model stays, that could be a really fundamental rethinking of the advertising. Uh, I was just told that OpenAI says that it has about 20 million subscribers, and that is a tiny drip in the bucket of a trillion-dollar buildout for data centers. So, that can’t be their model going forward. So, where will our access come from? Will it be like a healthcare model where you get access to compute through your job? And if so, that brings in all these privacy questions. But ultimately, I think there’s gonna be very different economic models for the AI platforms because I think they’ve learned this lesson around around advertising. I mean, it is interesting, though, because they have to pay for these extraordinary buildouts, right? So you’re having hundreds of billions, maybe trillions of dollars of infrastructure buildouts for companies that, yeah, are making a lot more money, but nothing close to what they’re spending on chips and infrastructure. And, you know, you mentioned OpenAI, and there’s this question of like, essentially, will you get tiers, right? Like, essentially, you can get some free tier, but it might have ads. Or you can pay a little bit, which is where you get the same thing but without the ads, or you can pay more, in which case you get more tokens to do using more compute resources, or you’re an enterprise and it’s even more. So there’s like whether there’s a tiered system. Anthropic had a Super Bowl ad making fun of that model. Like, they’re like, I think that we don’t know where it’s going to be yet. Yes, sir. And there’s a microphone right there. My question to Andy, which is about AI and security issue. So one particular security issue is arm race. So in the current cyber wars, it’s clear that AI is terribly important. So nowadays when people talk about AI race, race between China and the United States. So you didn’t talk about arm race. So my question is, to what degree it is aware in the White House or in policy circle and what should be done? That is also related to the issue of selling AI chips. To China, which is not only about quality, it’s also about quantity because the quantity is ammunition. I love that phrase, quantity is ammunition. It’s a great set of questions. You know, one of the contests is military diffusion, use of AI, and I think that the US has, I think, invested a lot of effort going back to even the first Trump administration and thinking through how the military can embrace AI, hopefully in a lawful manner. China has done the same thing. I have an acquaintance, Mike Brown, who who was a former head of the Defense Innovation Unit, the Pentagon’s kind of expeditionary Silicon Valley office, and who’s now, he’s a partner at Shield Capital, a VC firm. And he makes a point that the Iran War is the first AI war because if you look at metrics like the increase in the number of targets actioned over the first day or two, I mean, it’s orders of magnitude more than previous conflicts in Saudi Arabia. Because of the MAVEN smart systems. Because of the MAVEN smart, yeah, the AI-enabled MAVEN smart system. So it’s very much a real issue. You mentioned the export controls. That was the desire to constrain the People’s Liberation Army’s warfighting ability was the original rationale for export controls. In fact, you know, if you go back and look at the way that the Biden administration framed the chips export controls, was not about outcompeting China for AI leadership. It was fundamentally about the use by the PLA. Now that narrative began to change over the course of the administration, and it had the effect, I think, of muddying what had been decades of— and this is not necessarily a critique, it’s an observation— muddying a very, a pretty clean distinction in US export control policy between where export controls were never used for anything other than national security or human rights. And so this purity is no longer the case. And I think doctrinally, it’s much harder sell for allies to support export controls when you have these economic criteria factored in, because it becomes, you know, then not just about security, but about, okay, whose companies, you know, win or lose, you know, whether or not this export control goes into place or not. And the reason allies are important is, you know, there’s this saying in the export control world, you can’t dam half a river, right? You need to dam the whole river, and that’s where the allies come in to hopefully prevent that from happening. If the allies doubt your motivations, doubt your sincerity, then it’s much harder to sustain those coalitions over the long term. So I want to follow up just a little bit because, you know, Annie, in your opening salvo, you talked about this question about whether we’ll ever reach artificial general intelligence, what Dario calls a country of geniuses in a data center. Maybe, maybe not, but what we do know is that The models are getting superhuman at certain things, and all of the frontier labs have invested heavily in making them superhuman at coding because they all are making a bet on recursive self-improvement, that essentially the models get better at software engineering than, than the best humans, and they improve themselves, you know, and you get a flywheel effect. In fact, Jack Clark, who’s another co-founder of Anthropic, had a blog post yesterday or the day before Jack Moore, where he is predicting that by the end of 2028 you’ll have fully autonomous AI R&D. Who knows? But Jack’s a serious guy, by the way, and he’s not, he’s not prone to hype in my— and that’s just because the models are doing a lot more. And one of the things they’re getting better at is, is identifying vulnerabilities in software. And the argument for holding Mythos back for 6 months was to give us an opportunity to go shields up on critical infrastructure. Which is an area where we know China has implanted a lot of malware for the event of a US-China conflict. So I guess the question really, you work cybersecurity and policy issues at the White House for two presidents. If you were sitting there now, how freaked out would you be about Mythos, and how relieved would you be that, thank God, an American company developed it first rather than Deepseek or Zhiyi AI or Alibaba? Developing at first, because I suspect the Chinese Communist Party would not have put a pause on the— either the— well, they might have put a pause on it. They probably wouldn’t have told the world about it. So how freaked out would you be? I’d be pretty freaked out. I’d be pretty freaked out. And the reason I’d be freaked out is, maybe to just take a step back, what Mythos and actually in ChatGPT 5.5, which has not quite but similar levels of performance, Yes, they’re effective at finding vulnerabilities and the next step, developing exploits for those vulnerabilities. They’re good at that, as good as a pretty high-end human analyst. Where they really begin to shine is the ability to string together a multi-step attack, right? You know, when you try to infiltrate a system, it’s not as simple as, you know, You’ve got your vulnerability, you’ve identified your vulnerability, you’ve developed your exploit. Now you’ve got to figure out, OK, how do I get that implant onto the machine? How do I get to where I need to go in the network to extract the information that I want or to generate the effect that I want? And that’s a multi-step operation. And what Mythos and 5.5 were able to do is carry out those multi-step functions with very high degrees of success. And what that allows is the attacker to scale, not just identify more vulnerabilities and exploits, but to scale the attacks to take out a broader range of victims. Drew mentioned the biological threat actors or economic actors, so were cyber threat actors, right? And it effectively lowers the cost of carrying out more complicated attacks. Attacks. You know, there’s a debate in cybersecurity circles around, okay, to what extent will AI ultimately favor attackers versus defenders? And I think you can— the way I think about it is it’s a problem of dueling J-curves, right? So the J-curve describes the pattern that new technology typically takes when it’s being adopted in an economy. There’s an initial decline in performance as the users make investments that don’t yield real tangible results. They yield intangible results like more expertise, more knowledge. There may not be complementary investments in place yet to take full advantage of that innovation. And then performance picks up, right? That’s the J-curve. Electricity, you name it, semiconductors, GPS, they all kind of follow this J-curve pattern. With cyber, the attackers are gonna be able to flatten that J curve a lot faster than defenders. Reason being, they will be able to generate returns on their investment very, very, very quickly because they’ll be able to scale attacks in ways that make it very challenging for the defenders who have to have a much, I think, deeper curve to flatten. They have to figure out, okay, how do I use this tool? How do I incorporate it into my current network defenses? How do I use it in a way that’s not gonna cause unintended consequences? It’s a much more complicated, harder, expensive challenge to overcome. So I think for the short term, mythos is going to help attackers, especially go after less sophisticated, less resource, or more resource-constrained defenders. Jen, you mentioned before that Chinese authorities have been much more assertive in regulating AI. Licensing and, you know, they’ve also been, you know, we know that the CCP is pretty aggressive on both cyber offense and defense. Do you have any sense for how, I don’t know, the Chinese government is viewing this moment where it appears that American models are getting particularly good at cyber? Do you have any insights into that? I wouldn’t say that. I would say on the regulation front, we in the US tend to often equate more regulation as stifling innovation. I think on the Chinese side, what you see is very, very heavy regulation very quickly on political control, as related to political control, very little on kind of the privacy side that I think in the US we’re more concerned around. So I think we really, when thinking about, at least on the regulation side, need to think more in a more nuanced way of what kind of regulation different types might have different effects. But on the cyber side, I’m— yeah, don’t have any insights. Jim. Great. Thank you so much. Thanks to the panel. We’ve been talking about the federal government in multiple contexts here. Just curious, in the US context, the role of the states either in stimulating innovation or regulating it. I’m just curious if any of you wants to comment. Thanks. Yeah, Jeff, I don’t know if your research— I mean, I know you did You’ve done a lot of international, but is there, I have to imagine there’s high diversity at the state level too on social media. Clearly there’s, the Trump administration has expressed concern on the AI front that you have this patchwork of regulations that will stifle innovation and therefore you should have one model, but of course they can’t work with Congress to actually come up with that one model. But is there anything at the state or local level that— Yeah, so California is definitely a leader in the regulatory side, importance in since all of this is coming out of California. So, there is currently legislation in front of the Assembly on creating a social media ban and an equivalent eSafety to regulate it. And I don’t know, it has a decent chance, and that’s gonna have an influence on the rest of the states. On AI, there’s also a number of bills going through. Those ones seem to have less likelihood, like less likely pass forward. It does seem to be a bipartisan issue, both like states that are thinking about social media bans are red and blue. And same with AI. So I think— And that’s mostly because a lot of this is about protecting kids, right? Kids. Yeah, that seems to be one place kind of you can get everybody to agree. At the federal level, not even that. So there’s been a number of bills that have been in the works for almost a decade and haven’t moved through. So even though it’s kids, which is the focus, it seems to be states is where the action is. Interesting. Sir. Thank the panel for their excellent input. One area I think we’ve neglected somewhat is the medical area. If I ask the audience how many people use social media and platforms to get diagnosis and information, I’d probably get a pretty large handful of raised— of raised hands. 80% of physicians use AI-assisted programs for information. My concern is the misinformation, disinformation, politicalized information. People are dying because they didn’t get their kids vaccinated for measles. People are not getting HPV vaccines and going to get exposed to 7 different cancers. And so I think we need to address world-changing technology in respect to how we’re going to regulate the disinformation and misinformation from a medical point of view? Maybe we go back to— because, Jeff, you had mentioned this conspiracy theory. Yeah. So, you know, my sense is, first of all, the LLMs hallucinate a lot less than they used to, in large part because they’re able to access the internet in real time. Yeah. So they search and they also go through more of a step-by-step process. Doesn’t mean they don’t make stuff up. They do. But if there’s, you know, a pretty good distribution distribution of the data in their training and they go on the internet, they don’t make as many mistakes as they used to. But can you talk about like, okay, you go on and you believe that like the COVID vaccine put microchips in all our brains, you know, or that vaccines don’t, you know, prevent measles or other things. What does the research tell us about using large language models for in the conspiracy theory context? Yeah, I’ll share two things real quick. So David Rand’s work was asking about this function on Twitter where you can say @Grok, is this true? And what was really fascinating there is Grok was accurate a lot, and it was really providing these users. And what was really important, and there’s no, you know, monopoly on misinformation, but Republicans often are exposed to more, especially in the health space. And Republicans frequently asked Grok to fact-check not only Democrats but also Republicans. Democrats really only fact-check Republicans. Republicans are fact-checking everybody. And Grok was accurate and influential in that person that whatever they encountered. So I think to your point, there is some potential optimism here that these models which have a lot of information and people seem to be able to listen to differently than he is, has some potential. So I think there’s potential. Yes, and just the other point I want to make is like when there’s a, you know, the administration is changing and when we have institutions that now are really under attack and, you know, these credibility issues, it’s really difficult to talk about what grand truth is. Before we’d say like just look at the CDC site, that’s the best information available, and now it’s very difficult to say that. So that’s a larger challenge. So if for folks who are interested in that subject, I have no vested interest if you’re interested in this, I just really like the podcast Hard Fork. If you don’t listen to it, it’s funny and informative. It’s about tech. The last episode, the second half of the episode was entirely about the medical use of AI. And it was, so for your entertainment later in the evening. But Drew, you wanted to, you got a two-finger. One thing really quick and then an anecdote. Fake bio has been around longer than fake news, right? It’s come back in a way, and it’s not good. What’s an example of fake bio? Well, we just heard. Yeah, like health and medical. So here’s an anecdote. I was at an event, and after the event, I got cornered by about 50 septuagenarians. And they come up to me. I’m the only bio person in the room. And they go, why are you saying vaccines are good? And I’m like, wait a minute. I didn’t even say the vaccine word the whole panel. But they wanted to talk about vaccines. Vaccines. And so I said, well, let’s talk about smallpox because I’m on the smallpox committee too for WHO. And let’s talk about the smallpox vaccine. They want to talk about RNA vaccines. They’re like, nope, we’re going to talk about pox virus. And they remembered that because they got the vaccine. And the truth about the first generation pox vaccine for human pox is not perfect. It has a casualty rate. But we used it because smallpox is a scourge, and it was a political decision. It’s like straight-up political decision. And that was a reset of the conversation. So now we were having a real conversation. And so I pivoted, and I went to US Thanksgiving dinner. How do you feel about US Thanksgiving dinner? And they said, well, we like it. But why are you talking about Thanksgiving? I was like, when we have Thanksgiving, my wife’s My parents come over, and we try and make the best meal we can with the best ingredients we can find. It’s a little bit stressful. It’s a little bit high stakes. And we’re yelling at each other in the kitchen, but it’s out of love. And we pull it off most of the time. And they go, that’s what Thanksgiving’s like for us. Yeah, we appreciate that. And I go, what if vaccines were like that? They go, what do you mean? I go, what if vaccines were like Thanksgiving dinner? They go, huh? Like, when you need a vaccine, you made it. In your kitchen for your loved ones with the best ingredients you could source. And you knew there was no Wi-Fi chip or Bluetooth chip in it because you didn’t put it in, right? Like, it’s like there’s a little bit of design thinking applied to vaccine, right? And there’s even vaccines that are forthcoming that are living skin creams that tickle the immune system with no needle, right? Works in mice. Mice have thinner skin, so we’ll see if it works in people. But they go, oh, that’s the type of vaccine we might want. You know, so what if for a reasonable fraction of our population, like I might wish to get my vaccine from Pfizer, but other people might wish to have a Freedom vaccine, you know, or something like that, that they roll themselves, right, in their kitchen, right? Because that’s the one they want. They want to have that optionality. Backing up and trying to land a broader point, we have to deal with citizenship and literacy and distributed capabilities. One of the things I like the most as an idea coming out of FSI is the work of Callie Chappell, who was a postdoc at CSAC, and she took the word library and mutated it so the first letter I becomes the letter A, and now you get this new word called labrary. Labrary. And that then leads to a new profession called the labrarian. And so imagine if we reran the Carnegie playbook from 1890 to 1928. And we, we took our community centers and our public libraries and we expanded them to have access to natural sciences and engineering capabilities. I’ll start with biology, right? We don’t have our Carnegie yet, but, but we could do this soon enough to matter and let people meet these emerging technologies and natural sciences where they are. So that they can learn for themselves what the opportunities are. They could be engaged with a local community member who’s the trusted steward of these powerful capabilities and so on, right? So I think you’re raising a profound point. The approach in my mind isn’t to simplistically confront fake news through the social platforms head-on, but to listen with empathy to understand what people are wishing for and then figure out how to renew an investment in who we are, you know, as a set of individuals comprising a culture and population. Education and trust. Yeah, yeah. By the way, I have a very— you’ll notice I have a bracelet. I’ve worn this every day since early 2021. It says, “I got vaccinated at Stanford Medicine.” And there’s a story behind that, which is that I had a stem cell transplant in early 2021 that reset my immune system to absolute day zero. Yeah. So the very first vaccine that my new body got was actually the COVID vaccine in the middle of the spike in the pandemic, if you remember, early 2021. And so I made a big bet on whether that was going to keep my body with no immune system alive. And it did. I also— that’s the good news. The bad news is I had to get all my baby vaccines again and that getting them as an adult, no bueno. That’s not fun at all. A lot, a lot of jabs, but also saved my life. Sir. I’d like to bring the focus squarely back onto policy, if I may, and specifically onto the issue of international governance. You know, it’s been— being D.C.-based, it’s been very clear to me that saying what I’ve just said has been met with eye-rolling for the last 18 months, but that is changing, clearly. I think the public dialogue is changing. I can think of two cases of influential public commentators in the last two weeks— Tom Friedman today in the New York Times, Sebastian Mallaby about two weeks ago— saying, you know, let’s forget this business of winning the AI race. I have to note that there were two people sitting on that stage just yesterday that said very emphatically, we’re in an AI race, we We have to win. But I think that is a little bit disappearing now. And from my perspective, we are back not to safety, which is an outward, but to the concepts that underlay 3 or so years ago the establishment of the AI safety institutes in the UK and here. So I’m thinking it’s great if this gets back on the agenda for Trump and Xi. It seems to me it probably will do. And it’s going to be great if there’s some signal given at the top level, but any of us who’ve been involved in international negotiation know that there’s going to be a massive amount of staff work that’s got to be done. There are going to be difficult decisions about who’s going to negotiate, where are they going to negotiate, who’s going to be involved, is it going to be bilateral, is it going to be multilateral, all of these questions. So I wonder whether anybody has any thoughts about that, maybe taking as the starting some kind of a signal coming out of the Trump-Xi meeting that, yes, okay, we can start thinking about this again. So, Andy, maybe starting with you, but I’d love to have other views on this. I mean, you started off by saying that the race analogy is a little complicated. I’ve always thought it was complicated, but to some degree inevitable. Yeah. Because the companies are racing against each other and they have a market logic to do that. That isn’t going to be wished away. And it does have profound geopolitical consequences, which means the United States and China are racing. So we can say we don’t like the race, but the countries and companies are going to race. The question is whether there are going to be any guardrails around it. And I do think that the point that I think you’ve reached an inflection point a bit, I think largely because of Mythos, frankly, where the Trump administration having basically discarded the voluntary safeguards put in place by the Biden administration and the Biden administration’s executive order are now tacking back in that direction by at least gesturing at the possibility that the US government, KC, not AC, would have early access to the models and in theory could tell people to hit the pause button if there were challenges. I guess the issue is, like, do you think we’re at this inflection point? Is it— do you really— I mean, we’ve also seen some personnel changes, right? David Sachs is no longer in the White House full-time. Well, I don’t know if he was ever full-time technically, but like, do you think that there is space for this AI safety or AI security governance conversation to come back in an administration that was really skeptical of it? I think Mythos scared him, to be honest. I think, you know, this— there’s been a lot of news reporting about this in the last couple of days of a reconsideration of, of, you know, a ban. Not, not, not reinstating the Biden-era measures, but, you know, taking some ideas about testing models, um, red teaming models, and, and potentially again putting a pause on the release and reviving those concepts. Um, I, I think it’s— I, I think if I’m betting, I think there’s a pretty good chance that we’ll see some movement on, on that front. The The multilateral or even bilateral cooperation with China is a trickier gambit because if you think— take Mythos as an example, right? The administration hasn’t really been that actively engaged in— I mean, Glasswing, the project that Anthropic has pulled together to give a select group of companies early access to the model, the government’s not organizing that. In fact, the government has sort of argued over initially over who would sort of run point point on that initiative. I know that many foreign governments are frustrated because they have asked Anthropic for access to the model because they obviously have a huge equity at stake. And Anthropic has said no. I think with the partial exception of the UK, maybe. The UK, I think, yeah, that may be right. They’ve said no, you need to talk to the US government to mediate because we’re concerned about export controls. So, and then on the Chinese side, China has a law that requires that any vulnerability discovered in software first be shared with the government before disclosure. And the evidence is, it’s hard to gather evidence on what scale of vulnerabilities that the Chinese government is sitting on, right? Vulnerabilities has received through this legal process and has told the vendor or the researcher who discovered it, no, you cannot tell the vendor. No, vendor, you cannot patch this. But I have to imagine it’s substantial. And it’s difficult for me to imagine— I would not trust, if I was on the US side, given that law, how the Chinese could make any commitments to do anything, hold a Chinese company to anything like the standard that Anthropic is holding itself to. Yeah, I think the other challenge that you have to, I mean, the two other challenges are you had a hollowing out of a lot of the parts of our government that were trying to build up expertise on this. So even if you flipped a switch, it doesn’t mean that you have all the people in place at KC, which was AC. Struggled to have a new director. They hired somebody from Anthropic, actually, and then they— and that guy got lumared not too long ago. So that’s a problem, is whether you’re going to have the right people in the right place. The other problem is that, you know, we have something like 19 intelligence agencies in the United States. All but a handful report to the Secretary of War. And the Secretary of War has PNG’d Anthropic from DOD systems. So it’s also complicated for our intelligence community to get their arms wrapped around in a world where Anthropic has been labeled a supply chain risk that shouldn’t interact with the Pentagon at a time when they have the most advanced models in the world. So my suspicion is we’ll get into a better place, but we’ll see. All right, I think we have Jen, and then we’ll have time for one last question. Jen, please. On the bilateral side, it’s not clear to me what Mythos in terms of incentives for the China side creates. On one hand, maybe it creates more incentives for cooperation because they know they’re behind. But on the other hand, I think it really reinforces this view that China has had for the past several years that AGI is a problem for the US and the negative consequences of generative AI will be something the US has to deal with. For China, it’s going to be a boon. The economy. And I think this kind of optimism-pessimism view should, should be out there. And when you go to China, you talk to whether it’s ordinary citizens, policymakers, or academics about GenAI, it’s overwhelmingly positive. It’s about the possibilities of this new technology in improving people’s lives, the economy. Whereas in the US, when we talk about it, it’s all about the risks. Maybe that’s a little bit extreme. Extreme, but I think that sort of difference in overarching perspective potentially could have implications for how these technologies— it’s a super insightful point because in, in the Track 2 conversations that we’ve engaged in with our Chinese counterparts, actually the, the way in which the term artificial general intelligence is discussed in the Valley, let alone something like artificial superintelligence, is just not the way it’s discussed in China. Largely in China, it’s about applications. And there’s a belief, both an enthusiasm around it, and a belief that China is well positioned to develop killer apps, even if they’re slightly behind in the frontier. So it’s just a different— I think your point is super insightful. It’s just a different conversation. Totally. And so I think what will be interesting to see out of the summit between Trump and Xi is whether actually there was some psychic shift, because the Chinese Communist Party for the first time has started to talk about AGI in the way that Americans have. Have. And we’ll see whether this is a wake-up— we’re about to run an actual experiment, basically, I guess. All right. We got time for one last question from the audience. Sir, step up to the mic. Ask your question. Yeah, I was curious to hear what you guys feel about US and China’s propensity to put AI in like a strategic decision-making role internationally for international policy. So like right now where you have AI like enabling experts in one domain. So like imagine a world where we have one model being fed data feeds from intelligence, cyber world, digital domain, physical domain, public sentiment, and actually either in a conservative space providing recommendations on how to act in political international power struggles, but in the extreme case, actually actuating those digital and physical actions to tilt the scales in one country’s favor. I can’t speak to actually the foreign policy side, but domestically you see a huge appetite for already implementing these applications in judicial decision-making, in healthcare decision-making. So one of my students last year, Stanford student studying abroad in Beijing, broke her wrist in May, went to a Chinese hospital, got CT and X-ray scans before seeing a doctor. She already got an AI readout. So that means generative AI was already integrated into routine hospital procedures as of May of 2025. So the adoption rate and the tolerance for the negative consequences domestically are much, much higher than you see in the US. I don’t know if that will translate to foreign decision, international relations decision-making, but that’s what we observe domestically. But Andy mentioned, you know, so in the US context, in, you know, arguably the most strategic set of decisions a president can make is about war. And Mike Brown is brilliant. The MAVEN smart system was used in Ukraine before it was used in Iran. So maybe Ukraine was the first AI war. Who knows? But the point being that what the system does is integrate intelligence across domains to provide exquisite situational awareness of the battlefield, right, where they are, where we are, what they intend to do, It essentially generates targets. It then creates a recommendation engine about, given your priorities, which targets should you rack and stack and how should you service those targets, right? Which aircraft, what tankers from what bases? Now, these are all just recommendations, right? This is AI providing decision support. I think that we should expect that within the next 10 years, presidents of the United States will have an AI advisor not unlike they have human advisors, not in replacement of, because I think ultimately people are going to want to turn to human beings for these types of huge consequences. But I think you’re going to get a lot more advice coming from much more powerful models. And then the question, questions about how much cognitive offloading are you doing on these models? How much automation bias do you have? They’re also spitting out recommendations at a rate that humans don’t process quickly. And the closer you the more pressure there is to delegate. So these are all huge questions, but I don’t think that, I mean, I don’t know, if you’ve seen, you know, what’s your sense? Yeah, I mean, I have a project underway, it’s called the Future Decision Making, it’s aimed at trying to answer some of these questions. And I don’t wanna get too far ahead of where we are with our research, But the challenge becomes when it’s— OK, it’s hard enough when it’s just, say, just the United States who has this capability. But when an adversary has the same capability, it accelerates the need to make decisions, right? And then you run into questions about how do you even preserve civilian control, human control, when the machines themselves— and there are strong incentives to allow those machines themselves to operate at machine speed. And that’s a dynamic that, you know, we have— we’ve had the same, you know, Napoleonic command structure in the military now for, you know, for, you know, since the 19th century, right? And I think it’s an important question to ask, you know, how amenable that command structure is to this new mode of decision-making. We’re at time. I think one of the things that is extraordinary about this panel is that it illustrated the fact that these world-changing technologies are not happening in isolation of one another. I think especially at the intersection of bio and AI, that was evidenced here. But we could have had experts on quantum or space or clean energy and had similar intersectionality, I think, on display. So I think that is one thing that I took from this panel that I think is super interesting. And the other is it’s world-changing in every way in the in the sense that it’s going to affect relationships between countries, not the least of which between the two most powerful countries in the world. And we’ve talked a lot about US-China today. But also it’s going to impact at lower levels of analysis, all the way to, you know, the overall impacts on societies within countries and on individuals and their health and well-being and the opportunities that they have. And what’s so exciting is that FSI is a place where we can have all those conversations at the same time. So please join me in a round of applause for our brilliant panelists. And if you like what you heard, please come to our events in the future. If you don’t like what you heard, don’t tell anybody about it, but have a couple glasses of wine and maybe you’ll have an even better opinion. Have a wonderful evening, and it’s Been really great to be with all of you. Thank you.