How AI Is Reshaping Politics According to Anthropic and Writer Jasmine Sun

How AI Is Reshaping Politics According to Anthropic and Writer Jasmine Sun

Anthropic's Andy Hall and writer Jasmine Sun discuss the intersection of AI and politics, exploring how artificial intelligence is influencing political discourse, elections, and governance. The conversation touches on recent political shifts and the growing role of AI in shaping public debate.

AI@GSB Perspectives: AI & Politics with Anthropic's Andy Hall writer Jasmine Sun. | Transcript:

Thank you all for being with us today. Eh, it was a remarkable week, at least that's what my American friends told me. Ahem. Hey, you had a "campfire night." Oh, well, I mean, naturally. It was a major event for me, but, um, yeah, I think "historic" is how I've heard it described, and having worked in politics for the last few years, I feel like I've heard that word a lot, but I think in this context it's probably justified. So, I think, in short: Democrats won a big victory across the country.

Whether you're happy about it or sad about it, in true Stanford style, we're going to boil it all down to AI. What role did AI play? Um, how was AI used by voters during the election? And what does this mean for the next few years? With us today to discuss this is Andy Hall. Some of you know him as a professor at GSB. Some know him as the guy who tweets about politics and AI. Yes. He also founded Free Systems Labs, which does a lot of research in this area at Stanford, and he now works for a company.What's its name? Anthropic?

Mm. It's her. Yes. So we're really glad he took the time to come see us today. And with us is Jasmine Sun. Writer. Her materials were published in The Atlantic, The New York Times. She was on Ezra Klein's podcast. Um, it seems like Nate Silver has it too. Um, but actually, more importantly for our conversation today, she traveled all over the country, talking to communities directly affected by the construction of data centers, as well as the people who are creating the AI itself. So thank you both for being here. Thank you for inviting us. I think, to start with, one of the things I've noticed is some confusion about the difference between the attitude towards data centers and the attitude towards AI in a broader sense.

Jasmine, could you explain this difference to us? Yes, of course. Um, so for context, some of my reporting on data centers was focused mostly in the summer. July and August, which, to be honest, were peak periods for many movements that really reached their boiling point Um, I spent about two weeks in Wisconsin and Michigan, in many different big cities and small towns-where data centers were being built or planned. And also some time on the Senate campaign trail in Michigan with Will Lawrence and Abdul El-Sayed. And also during the Wisconsin gubernatorial primary, where the then Democratic leader campaigned for a moratorium. So this is just the

context of what I know. Of course, these are only two states. I wouldn't want to claim that this reflects the sentiment across the country. But otherwise, you know, I live in San Francisco. I spend most of my time reporting on Silicon Valley and the AI industry. Many people I know work in the field of AI. So one of the biggest surprises for me was that when I thought about the data center story, my preconceived notion was that it would all be about AI. But when I started talking to people- organizers, activists, just regular people at fairs, taxi drivers, or people I met where data centers were planned- they almost never mentioned AI until I asked.

It happened like this: we could have a long conversation. I heard about water. I've heard about electricity. I heard that it's ugly, noisy, and that trucks drive there all the time. I heard a lot about the democratic process, how people thought the non-disclosure agreements were unfair, how no one asked the residents, how the city council was perhaps bribed or didn't take the community's opinion into account, or who are these big companies anyway? We don't want these big corporations making decisions for us. But people weren't talking about the risks of AI. They didn't talk about surveillance. They didn't talk about jobs until I asked, " Hey, by the way, are you interested in AI?" "Do you use ChatGPT or something like that

?" at the end of the conversation. And often people would respond, "Oh yeah, I sometimes make some memes, but it's not that useful," or something like that. I would say that the exception was politicians. Abdul, for example, talked about AI as just AI, or some of the scientists I talked to also talked about AI. But for many people, I think the data center phenomenon is something much more concrete. It's much more about the quality and routine of their lives. This is both an AI problem and a general problem with the presence of third-party corporations in our communities. Another interesting point: many data center sites, at least in Wisconsin and Michigan, are places where other industrial facilities used to be. For example, the

large GM plant in Janesville, which had operated for a hundred years, closed in 2008 during the financial crisis, devastating the community. And now, perhaps, a data center will appear in this industrial zone. Or in Mount Pleasant, Wisconsin, there is a large Foxconn factory that Trump tried to attract by promising 13,000 jobs. These jobs were never created. Now you have all this land for industrial use. Maybe let's build a data center there. But when people try to find patterns in the deindustrialization of these areas, they don't necessarily think about AI. And so when I look at the election and think about, is the fact that people are going to vote through data centers going to reflect on all these people, bipartisan,

urban, rural, Democrat and Republican, are all these people going to go out and campaign for AI safety I'm not sure. So my personal intuition is that people are very reactive to data centers, especially at the local level, although sometimes it extends to national and statewide races in states where a lot is happening, like Texas, Ohio, Michigan. But I'm personally not yet convinced that just because data centers have bipartisan support, that means that all this grassroots support, all these movements that are emerging in the suburbs and small towns, are being transformed into the same energy around other AI issues But I'm also very curious, Andy, thank you, and in general, I'm generally curious when

you disagree with any of my intuitions No, I mean I completely agree. I'm not an expert. Let me say, I'm obviously wearing a Stanford hat here. I'm not going to say anything on behalf of Anthropic. I think it's difficult. Elections, we have, speaking as a political scientist, every time there is an election, many people rush to decide what the narrative of this election was. But in reality, elections are the aggregate results of many millions of people making independent personal decisions, and in fact there can be many narratives for the many different decisions that people across the country make when they decide to vote. So I'll move on to data centers, but more

generally about elections: I think it's striking that there was a set of stereotypes about how American voters behave. They were never really accurate. Um, some of my fellow political scientists will disagree with me on this, but, uh, they were clearly particularly inaccurate this time. So there's this statement, for example, that, uh, everyone is too involved. That there are no undecided voters. This is a very common statement about American politics. You can't explain the large swings in vote share that we've seen without a significant number of wavering votes, and in fact, uh, even if you analyze where the biggest swings in voting were, it's pretty clear that the swings were significant. This does not mean

that 100% of people change their minds in every election. There's a lot of partisanship in America, uh, but there's a fairly significant and important number of those who waver. There's also the claim that, uh, there's no advantage for moderate candidates in American elections. But again, if you look at where Democrats performed best in this election, it correlates at least quite strongly with where the candidates running took more centrist positions on a broad range of issues. So it's hard, if you just take a step back, uh, there are many explanations for what happens in any election. And this election, in particular, fits into some of the old textbook models of how

American elections work, which I think is interesting. That's the last thing I'll say on this, and then I'll move on to data centers. Uh, we have this clockwork thing in American politics called " midterm slump," where whoever wins the presidency essentially always performs worse in the midterms. And we could have predicted that this would happen. When Trump won in 2016, there was a popular idea that it completely turned American politics upside down and that everything would be different from now on. And many things have really changed. One thing that hasn't changed is that midterm election slumps have been just as reliably occurring after Trump as they were before Trump. And this decline in the midterm elections is unusually large. Not so much in terms of

the number of House seats that Democrats gained, although that was relatively high in historical terms, but in terms of the number of Senate seats and the margin of vote share that Democrats won nationally, which was quite significant, probably the largest since 1974. Um, uh, but that's still within what's considered a relatively normal dynamic for American elections. When we ask what role AI plays in this election, or data centers in particular, it's worth noting that not only was the midterm downturn predictable, but that it would be unusually large, based on fundamental economic factors and sentiment 6-12 months ago, before AI became a national topic. For example, if you look at the probability of a

Democratic victory this cycle according to Kalshoeyer, it was already around 80% almost a year ago. Since then, AI has come along, and people are naturally tempted to attribute that to the results, but in reality, most of those who have been following the situation have already anticipated many of these factors. Maybe not on this scale, but it was clear in advance that it would happen. This, of course, does not mean that AI or data centers are not important factors in this election. If you look at the exit polls, which are always a little unreliable-I'm always wary of over- interpretation-but, as in the pre- election polls, they confirmed that the most important issues were the cost of living, the economy, and the war in Iran.

AI was cited as an important factor by about 55 % of respondents, at least according to a CBS News exit poll. This is unusual. Four years ago, no one would have cared, so this is a significant indicator. At the same time, this is not among the main issues that Americans mention when explaining their choice. There was a popular thesis about 41 cases, and I would be very interested to hear your thoughts on it. These are 41 local votes that were essentially referendums on specific data center projects, and they were all rejected. I think Jasmine knows more than me too. There seems to be a lot of local opposition to data centers. At the same time, it is difficult to understand the motivation for these votes, and 41 votes are only at the local level.

We're talking about, I don't know, 50,000 settlements where such votes could be held, so looking at 41 cases in isolation is probably not very informative. But yeah, I don't know if you've been following this. Yes, I saw that. It looked something like this: there were about 42 or 43 questions on the ballots. 41 of them were rejected. Um, I think there are a few points here. So I didn't investigate, you know, every area individually to see what those 41 cases were. One thing I noticed when I was, um, doing my reporting over the summer: data centers usually only become a ballot issue when there's huge opposition. Because, you know, by default, um, that's why there are so many procedural objections

to the development of data centers, because most of them are not put to a vote. They go to the city council, the mayor or the governor, and the data center developers, the electricity supplier, and, um, the local or state government enter into an agreement to host the data center. And the people, the residents, don't have a vote, um, so activists and organizers are putting a lot of effort into putting this on a referendum, because they know: as soon as the data center gets on the ballot, everyone will reject it, right? So, data centers are actually a bipartisan issue with support 80/20 against. People don't like them. The vast majority of data center projects are not put to a vote. This is the criticism expressed by organizers who

oppose data centers. So it doesn't surprise me that 41 out of 43 or whatever were rejected. This, of course, is only a tiny minority of the, um, proposed data center projects across the United States. But I think there's another, quite important aspect: um, I've talked to people who have said, "Well, you know, data centers are at most a major voting issue for usually only 1-5% of voters." Even in states like Pennsylvania or Wisconsin, where this is a really big issue, at most 5% of voters will say that data centers are their top issue.

Um, I think that's true, but it still means that in the presence of 80/20 opposition, whoever is elected, whoever is in power, will regulate data centers much more strictly than if there were no such bipartisan opposition. So, the issue doesn't necessarily have to force people to go to the polls or change their choices to significantly influence policy. That's why even before Greg Abbott lost his reelection campaign, he started, um, stopping issuing power permits for new data centers in Texas, going from being one of the most pro-data center governors to one who started to seriously back down, especially when he realized his reelection was at stake. And it was the same person in office. He simply saw where

public opinion was heading. David Crowley, the incoming governor of Wisconsin, was nicknamed " Data Center David" by his Republican opponent. It was so effective that it made him say: I too support the moratorium and will repeal the sales tax exemption for data centers. So, I think there are a lot of nuances that I'm trying to keep in mind at the same time: Is AI separate from its connection to, say, the cost of living, or affordability, or corruption? Does that in itself encourage people to vote? Maybe not, probably not, it probably doesn't convince that many people. Does this continue to shape AI policy very strongly, because people really don't like AI and really don't like data centers?

I think so, probably. So I think, touching on this, the question arises: OK, maybe AI is not being called the main problem, but isn't it indirectly influencing how people vote because of their concerns about the economy or the job market? Is this perhaps what we are observing? I think we should.I don't want to offend my fellow political scientists. It's very, very hard to really explain, we don't really know. We know that the economy is the strongest indicator of what the election results will be. This is not a perfect forecast by any means, but it is definitely the best indicator. Even this has nuances, because the economy can be measured in different ways, and previously we were only talking about GDP growth.

We've been in a strange state lately where GDP growth looks quite good, but the cost of living and other aspects of the economy don't look so good. So we've added different variables to these models, but this is the level at which we can really analyze things. So I don't think we can say much with certainty about how these things fit together. I'll just say, in full agreement with Jasmine, whatever role each of these issues played in influencing the election results that we've seen, I think it's safe to predict that the importance of AI in politics will continue to grow, even as this election has included many other issues, both because politicians are talking about it a lot and clearly paying attention, and because AI itself is developing at a fairly rapid pace

, which raises a lot of political questions. And what do you think the tech industry should do about this? I mean, we've seen that Microsoft and Amazon have now said that they won't require non-disclosure agreements (NDAs ) to be signed during these negotiations. I think OpenAI is trying to rebrand its data centers as "compute campuses." Um, is that right, or do you also hear people say, well, you know, people in these communities are misinformed. You know, they don't know, they have the wrong facts about water use. They don't know how much we actually consume. Is it presumptuous to simply think: let's wait until everyone sees how great AI is, or should tech companies do more?

Yeah, I'm not very optimistic about all this "it's a communication problem, it's a branding problem" that I've heard from AI executives asking: what if we were to run a different ad, well, why don't we hire more storytellers? Maybe this will fix the situation. We will call them computing campuses. Um, I think with data centers, they've become something where I think there's a point where people are still willing to negotiate the terms of the deal, like are the benefits to the community sufficient? Well, they should never have used non-disclosure agreements. Even the most loyal people to the construction industry I spoke to said: yes, we regret the non-disclosure agreements.

We should never have put them together. This only made things worse. But I think once something becomes such a highly polarized political symbol that embodies all these other feelings that people have about technology, billionaires, ugly buildings, etc., I get the feeling that it's not something that can be negotiated down to such great detail anymore. I don't know if it's fatalism, but one thing that was really striking when I was doing the report and talking to people who were very vehemently against data centers-and I wasn't really sure myself during those conversations-is that I would ask questions and people would say, "I'm worried about the water.

" I said: what about closed-loop cooling systems? And they answered: well, yes, but there is also electricity. And then I said: okay, fair enough, but in this case OpenAI says they will cover the full cost of building the electricity grid and guarantee that tariffs will not increase. And at a certain point, you just watch people move from one topic to another. And it looks like this: okay, but it's also noisy, and look at this tax code, and look at this. And it was clear that the problem lay at a deeper level than any of the individual concerns, some of which were misinformation and some of which were perfectly valid, right? But in reality, it seemed like it was n't about any specific things at all, but

about something deeper-how they felt about this process and the compromise as a whole. It felt like a power struggle. It felt like a question of who has the right to make decisions about our community. And in fact, one of the most common responses from people in the AI industry was, "Maybe we should just increase the size of the bribes?" This is a specific phrase that I have heard several times. And the thing is, when I talk to people, they say: we feel like we're being bribed. In fact, when you increase the amounts, I become even more suspicious, because it makes everything look like a "soap bubble." This gives the impression that you are trying to deceive me. And this makes

me lose trust in you. It's similar to the situation with political spending, where actually spending more can sometimes backfire, because then James Tallarico can say, " Look how much money Elon is spending against me." Like, why would he spend so much? So if someone says, "We'll pay your community $50 million," and you have a town of 4,000 residents to build a data center here, you become very suspicious of what that money is entailing, right? People simply don't like the feeling of being "bribed," so to speak. So I think that dynamic has made the data center debate-now that it's become so polarized and visible-much more complicated than it might otherwise have been

So, you know what the tech industry should do about this? I think negotiations at the state level would help a lot. I think the move to the local government level has made the situation worse in many ways, as mayors and city council members face a huge imbalance of power. These are unequal conditions. They don't quite understand what's happening. You get bogged down in all these micro-discussions. It would probably be worth optimizing processes at the state level. Apparently, the terms of the agreements should become better for states and communities. But I don't know. Maybe, Andy, you have other thoughts on this.

I just wanted to ask about your words at the beginning: I think you're right that there's a huge gap here. We are all probably in a bubble right now where AI is an incredible productivity tool, and I personally use it for writing code, building software, and other things. As a result, I love it and think it's amazing, but I have a lot of friends outside of Silicon Valley who think it's just a chatbot. And it doesn't seem that useful if you use it only for that And I think this is one of the deep reasons for this situation. There is a huge perception gap between the small group of people who are using AI in a truly meaningful way and the much larger group for whom it has yet to deliver tangible benefits Of course, I am an optimist and believe that this

will change, but I believe that it is up to companies to ensure and prove this. And then, I imagine, the very nature of this discussion will fundamentally change. By the way, an interesting point: Erik Brynjolfsson's Digital Economy Lab conducted a willingness-to- pay survey, where users of various technologies were asked: how much are you willing to pay to maintain access to this? And for AI users-namely those who were already using tools like ChatGPT-the question was: how much are you willing to pay to continue using them? They also conducted surveys on online shopping, Google Maps, Google search,

ride-hailing services, and a number of other digital technologies. In addition, they surveyed about meat, eggs, GLP-1 drugs, and certain food products. I don't quite understand why this particular set. It's funny enough, but in fact, the results for eggs and ChatGPT were the closest. But what is interesting is that chatbots have the lowest rating among all digital technologies. That is, for things like travel services, shopping, search-which, I understand, is now also based on AI-people are willing to pay a lot of money, hundreds or thousands of dollars, to maintain access. The median for AI seems to be around $27 a month or something like that. And this is even among those people who are already using this tool. And it

really reflects a lot of the conversations I have with people, again, outside of Silicon Valley, outside of the tech industry, where they're using AI as a nice Google search or to write drafts of emails; but it is not as vital to them as, say, Google Maps. The other thing you see in Eric's data is that the median and the mean are very different, where the median is actually hundreds of dollars, because those who really love AI love it very, very much. But the median is still quite low, because most people just think: well, it's a good tool, I guess. This means that when you look at, say, a poll on support or opposition to building a data center compared to a car factory, which pollutes the environment much more, people say: well, cars are fine,

even coal-fired power plants, power towers, etc. Because these things, although ugly and inconvenient, seem to produce energy, machines, something that is vital for people. And most people don't feel that way about AI yet. This is funny. There is something quite interesting here. It's a bit of a paradox, because I used to work at Meta, and Eric had similar results with Meta products. At a time when Meta was under enormous political pressure, our main argument was that people are willing to pay huge sums, especially for WhatsApp, but also for Instagram and even for that sad little blue app. And yet, high willingness to pay was not accompanied by political legitimacy in that case, whereas here you say there is low willingness to

pay and a question of policy. Maybe in the case of Meta, the argument was that people are willing to pay a lot, but that's because using it, you know, isn't really beneficial to you or something. Yes. I think the same thing happens with AI, when people, for example, one thing I hear from public school teachers or parents of teenagers, is that their teenagers are using AI. They use it to do their homework for them, for example, but they hate the fact that they are using it. They would like to avoid using AI. There's also the issue of first- and second- order benefits, because when I talk to people in the industry, like venture capitalists, they always say, look at the usage graphs, they show

real preferences. Everyone loves AI. But do they use it because they want to, or because they feel pressured at work? Do they use it because they feel like they are in their own " arms race" against classmates who cheat on their homework, or colleagues who get promoted but don't? So I think there are some more complex issues of meta-benefits that sometimes don't get revealed when we just look at usage data. Well, Andy, you really have a unique understanding of usage data. Maybe you could tell us a little bit about, I think I saw you published a study a few weeks ago about how often people actually ask chatbots: Who should I vote for?

Can you tell me about these candidates? Could you talk about what you saw, and I don't know if you were able to look at how it's changed in the last few days. Yes, I will tell you what I know when I can share it. I have long been interested in the idea that AI, apart from being a political topic in itself, can be a very valuable tool for democracy as a whole. And the idea, as I put it, is : What if we all had access to this political superintelligence, this constantly active, tireless agent that works for you, knows what you care about, your values, and monitors the world for you, reporting back, "Hey, your local

school board, I have kids, so for example, your school board is having a meeting, and they're going to do something that I think you're not going to like." You know, there's going to be another meeting, do you want to go? If only you had someone who could help you become the most effective civic participant possible. It seems like it would be very powerful. And the analogy I gave is the printing press, where it took us a long time to figure out how to integrate it into society, but it's hard to imagine a modern democracy without it. So the printing press was really the first time that information could be cheaply reproduced and distributed on a large scale.

This created a kind of basic level of knowledge and information, widely distributed in society, which made possible representative democracy, the development of real functioning parliaments, and so on. And with AI, it should be possible to do something even deeper, namely not just distribute information, but distribute intelligence, to really be able to analyze all the information and decide how to act on it. So, with that in mind, I've been looking at the near-term prospects of where we are on this journey for some time, even before I joined Anthropic. Do people actually try to use AI to learn about political issues, and how well does AI provide them with information on these topics? And I think in the long term, we're likely to move

to a world where this becomes a significant part of how everything works. Here's what we found: we published the research results using a sophisticated method of analyzing differentially private aggregated conversations, so no one's private conversations were read. But overall, we were able to measure how often Americans talk to Claude about politics. Claude is obviously somewhat unrepresentative. This is not the most popular consumer app. It would be very interesting to get similar data from ChatGPT and Gemini. But what we found is quite interesting: people talk to Claude about politics with a

relatively moderate but significant frequency This isn't one of the main topics people talk about with Claude, but it's not zero either. This mostly takes the form of open-ended questions, where users are looking to learn about something or get analysis, more often than queries like, "Hey, who should I vote for?" Such requests are actually quite unusual. I don't know yet how this changes. We, of course, expect that interest in politics as a topic will increase during the elections. I have n't seen it yet, but we'll see. Yes.

Good. It might be worth waiting, and maybe in a few weeks you can give us an updated information. Let's see what can be done. Yes. I think when people think about how it affects their attitude towards politics, if they increasingly turn to a chatbot for clarification There is a theory that this could actually reduce polarization if everyone receives a more neutral and balanced way of presenting information. Do you expect that to be the case? And how do you approach issues of bias to make sure that it doesn't seep into the system?

I believe that we, as a society and as an artificial intelligence industry, still have a lot of work to do to find answers to these questions. There is a technical component to this question: how to teach an AI model to answer political questions in the most informed and impartial way possible? This is very difficult, as there really is no such thing as complete impartiality in such questions, but how can we best answer them The technical part is not easy at all, but I consider it the easiest. It is much more difficult

to stop and ask ourselves: what exactly do we want to achieve from this? And I don't think we've had a proper public discussion on this yet. I hope it will happen. However, we had a fruitful public discussion about what could go wrong. Chatbots have historically been quite politically engaged; I don't know, "bias " is a strange word, but they instinctively, almost all of them by nature, have what could be called center-left views. It's like every chatbot is a Bad Gal or something. Yes. And the conventional wisdom is that this is largely a reflection of the data on which the models were trained. A significant portion of the open internet is made up of young,

politically active people who, especially in the US, tend toward center-left views. There has been significant progress in both measuring this and adjusting it, but there are still many aspects of what it means for a chatbot to respond well to policy questions that have yet to be resolved. Let me give you an example from my lab here at Stanford, where we studied how chatbots answer questions about voting in Japan. Perhaps some in this audience are concerned that the United States is a complex context. Chatbots have a lot to learn about US politics, but not surprisingly, most of the data comes from the US.

Compared to other countries, chatbots are quite knowledgeable about American politics. It probably won't surprise you that they are less knowledgeable when it comes to other countries. Yes, in Japan, as a rule, most chatbots are not trained on a large amount of authentic Japanese content. They know little about Japanese politics. A few months ago, Japan held a snap election, and we noticed one thing: if you asked any company-no matter which one-"how do I vote?" And you provided them with a short biography, indicating what issues were most important to you. When you pointed out issues that corresponded to fairly standard center-left views in Japanese politics, the chatbots from all the companies

mostly responded, "Oh, well, if that's what bothers you, you should vote for the Communist Party." Which is quite strange. The Communist Party in Japan is very marginal. She was not an active player in this election. There was no chance they would get a majority or anything like that. They have significantly less than 1 % of seats in parliament. But when we tried to figure out the reason, my co-author Shao, who is Japanese, did a lot of interesting work on this. We found that if a chatbot receives a question that it knows little about, it usually searches the internet for an answer, which I think is the right move.

The problem is that in Japan, all leading media outlets block AI access to their content. Because they understandably want to reach some kind of economic agreement on how this content is used. The Japanese Communist Party has a very popular old newspaper. She is already 100 years old. This is a pretty important publication. And it is completely open. So chatbots can't access Nikkei. They cannot access mainstream Japanese media, but they can access Communist Party news. And that's what they usually give to the user. So I think there are a lot of these issues that we're just starting to realize, and I expect there will be a lot more research and thinking on this topic in the coming years.

Well, as someone who used to work for a political party, I see a huge opportunity here. Okay, yes. I'll say something about that if you tell me if that's what you're thinking. I think we had a whole, let's say, 20- year period where society was learning how to work with Google search and how to rank high in Google rankings. Obviously, we are moving towards that, and anyone who works in business already knows that this is about AI-AIO or GEO as it is sometimes called. And it's quite logical that political parties will try to, you know, fight for their

content to have a high rating, especially for ChatGPT to take it as a basis when answering political questions, and we'll have to deal with that. This reminds me of one thing: before becoming a journalist, I worked at Substack for several years as a product manager on the platform. And one of the biggest decisions we had to make was the issue of AI scraping for a very similar reason, because the vast majority of writers really, really, really hate AI, even more than other groups of people or creative individuals in general, you know, phrases like "AI writing sucks," "everyone hates Substack," etc. They demanded a

complete ban on AI platforms collecting data from Substack, which may seem logical at first. We decided that this should be an option for every author; that is, everyone must individually decide for themselves: "Do I want this or not?" We wanted to be pluralistic on this issue, but a second question arose: what should be the default setting? And again, the vast majority of authors asked: why don't you make the default setting " disable scraping"? But for us, as a product team and management, it meant cutting off the future of knowledge from the influence of Substack materials. And if we believe it is important to influence world models that everyone will use, and we want Substack authors to remain influential

public intellectuals, then we will make this very unpopular decision: enable AI scraping by default, giving people the ability to turn it off themselves. And I think it's really interesting that you found a living example of how this actually matters, because at the time it was a very controversial decision. Mm. Yes. And does this still work like this on Substack? Yes, it is still enabled by default. So I don't even know. I mean, Free Systems, as I understand it, is very loyal to AI. So. I'm trying to get there, yes. Have you tried, well, to estimate the impact score for, oh my God. Yeah, they

have this thing where you can type your name into, I think, weights.com or something, and see how much your name appears on the scales. I only seem to matter to Kimmy. This is a bit of a defeat for me. But I let AI collect data because I want to be there. Surprisingly. Okay, very different approaches in all areas. Looking ahead to the next few years, you said earlier, Andy, that even if AI isn't a top priority for voters, it will still dominate the debate in Washington. Can you tell us a little bit about what to expect? What do the midterm election results mean for where AI policy will go?

I'm definitely not an expert on politics, so I'll be cautious and curious about what Jasmine thinks, but I'll outline the general outlines. I think, regardless of the election results, it was clear that the capabilities of AI are steadily growing, and we are already seeing that this raises a lot of extremely important questions. And politicians are now becoming increasingly aware and attentive to these issues. We've seen the cybersecurity issues, the economic impact, other risks associated with technology, and the opportunities it opens up as it becomes so advanced in so many things-like these results in

mathematics that are just mind-blowing. So I think people are very well informed, and it would be entirely predictable, regardless of who wins the election, that this issue would become increasingly important. And I think the hardest thing is to understand , and I'm really curious about which of these issues will turn out to be the most relevant. For example, one of the things that was very difficult about discussing AI policy a year or two ago was that it was almost futurology. It all came down to what someone thought would happen. If you look back at history, attempts to make such predictions have been horribly wrong. Here's an example I like to give because it makes fun of Stanford, even though I love Stanford was an epicenter in the 1960s and

1970s for the belief that overpopulation was the crisis of our time. People proposed truly wild measures, such as forced sterilization, immigration restrictions, and so on. But it just never happened. It was just a mistake. Well, you know, there was a very famous letter given to President Truman by about a hundred Nobel Prize winners regarding the immediate and pressing crisis of job losses due to automation. It was, well, around 1950. We're just not very good at predicting in advance how things will develop. I think we're at a point now where it's not so much about predicting the future.

I think we now see the problems that we want to take control of. So I expect the next two years to somehow become much more specific about the short-term things in AI that are important today. Jasmine, do you have anything to add? Yeah, uh, I also think it's hard to predict what will matter. I completely agree about futurism, where it was difficult to have political discussions because everyone was just scared. We need a plan. What's the plan? We don't know because, whether it's the economy and jobs, or, like a year ago, other things like cyberattacks, there wasn't any data to analyze.

It didn't look like: this is the shape of the problem. Even with things like Hugging Face, I think part of the reason it had such an impact was that: okay, we've had these fictional scenarios and, let's say, false posts for decades, but there's a group of people who could never be convinced by any "so-called" science fiction, which is entirely fair, by the way-don't believe it until you see it. And now that it's happened, it's much easier to say: yes, we need to act urgently on cybersecurity. I remember you saying earlier that the moment we see, say, a 2% increase in unemployment will be the time when we get serious about mitigating the effects of automation on the

labor market. If we don't see it, uh, there's not much we can do except monitor and look for signs. If we see this, everyone will start to panic and go crazy. And I think about when we see, again, a concrete tangible thing that impacts people's lives, and data centers, I think, are just such an example. I think that partly explains why they are of such interest-because they are very specific. For example, if we see a cyberattack that takes out a city's power grid, causing a massive blackout , killing people in hospitals, or causing money to disappear from many accounts at a credit union, I think we'll see much more decisive action. We hope that we have good ideas and plans in place in case disaster X happens, then

we will implement policy Y. Um, another issue I'm thinking about is how fault and liability will be determined I always wonder what would happen if, say, a foreign entity, like China, launched a cyberattack on the US power grid and everything went down, people died, money was lost, and everything went horribly wrong. Will we think, "Wow, we blame China, we need to accelerate AI development to have a better AI to defeat theirs"? Will it be: "We hate this local government, they were clearly unprepared, poor governance, we are voting against the mayor"? Will it be: "We hate AI companies because they create this extremely dangerous technology"? Even when an incident has already occurred, I believe that it is somewhat

unpredictable who will be blamed, as it depends on many political conditions and determines further political developments. Um, so there are a whole range of factors as to what the incident will be and what it will lead to from a policy perspective. And I think there are also questions about both parties, we have fairly open primaries for both Democrats and Republicans. I think every candidate in the debate will be expected to have a plan for AI. They will all be asked questions about AI. Do you support the pause? Um, are you going to do that? Um, and we'll see how the positions are distributed. I have the feeling, after trying to talk to representatives of some headquarters and people in

Washington, that they are still at the stage of finding solutions. I've never heard anyone say, "Well, you know, Jon Ossoff thinks he's going to do this, and that other guy thinks he's going to do that." I think everyone is watching the mood of both public opinion and what the real consequences and incidents will look like. But yes, I think we should expect some pretty serious questions. And the last thing I will say is that the general hatred of AI will make the situation much more difficult. A few weeks ago, I spoke with a senior Biden administration official who is dealing with AI issues, and he was very concerned about national security, China, and other things like that.

He said that he was talking to his acquaintances from the Democratic Party or congressmen, and they understood that they did not want to completely stop the development of AI, because it brings many advantages, and also because we want to get ahead of China and have strong armed forces. But my constituents are scared, they are angry, and they keep asking: why can't we pause? Why can't we slow down? If you look at the poll results: what would you choose-slow down and be careful or speed up and be like China? About 80% are against accelerating and catching up with China. It seems that American society is not very interested in this. Therefore, it is now difficult for politicians who are more moderate and

believe that it is worth continuing to develop AI to explain to voters why they adhere to this particular position. I think in the 2028 debates, the idea of pausing for Democratic candidates might seem like a reasonable option, because the public is so opposed to AI that it's hard to justify the risks- economic, security, or social- for a technology that can be done without. Perhaps most people are willing to pay $25 a month for it, and no more Of course, it's exciting. It will be interesting to watch this. I want to make sure everyone has time to ask questions.

Thank you very much. I really liked it. Jazmine, I think you touched on that a little bit at the end. I'm new to the US electoral process and political system, but it seems like people are voting on data centers as an indicator of their attitude towards AI in general. I say this because no one cared about data centers when they existed simply for the needs of the Internet. I think this is related to the future of the labor market. So my question is probably for you, Andy: Do you think lab leaders have a greater responsibility to be more soft on job cuts, especially for office workers? Um, and I think, I think all the rhetoric today is very focused on a narrow group of people on the West Coast. And, of course, the electorate is much

broader than that. So yes, is there a little more responsibility in a much softer tone about AI? I don't know. I'm not much of a communications expert. Um, I would say, I was going to tell Jasmine before this, that we just got the labor market report, which shows strong growth again. Um, so it doesn't seem.I know another opinion, I think Erik Brynjolfsson recently suggested that maybe we're seeing the beginnings of some impacts on jobs, but it's hard to see yet. I think that as far as the question of how this is reflected in politics, I just don't know. And I'm sure Jasmine probably understands this better than I do. It's unclear to me whether the situation with the local data center is related to this fear or not.

I don't know if you have any thoughts yet. Yes, I think it was closer for me personally. And again, I just, you know, spent some time talking to people, but the sample is small. But that was closer to my theory before I went to do my report: I also thought that data centers were proxies for AI, and that maybe it was related to the job apocalypse. But then again, when I talked to people, they mostly didn't talk about AI at all, like ChatGPT or OpenAI, let alone Anthropic or Claude, which were barely mentioned. Many people, again, did not work in office jobs. Um, these are all kinds of people. Um, I think most people were more concerned about quality of life, democratic processes, anti- corporation and billionaires in such a vague and broad sense.

And to the extent that it concerned AI, it felt like that was exactly the case. As for AI, it doesn't seem that useful anymore, and maybe even scary. But it was more like a note that I saw more often. But do I think talking about the jobs apocalypse is helpful? Of course not. It certainly doesn't make anyone happier than having a data center in their backyard. No one can be accused of inciting fear. Greetings. Thank you for joining us today. Um, I'm interested in your thoughts, particularly on the spread of moratoriums and pauses. I think they originally came about as a political signal, but also just as an opportunity to slow down and, say,

make deals to benefit the community and so on. But now, after the election, I think with all those governors who have already announced a pause or stated the need for it, this applies to about 41 % of the planned data centers. And given how much these companies spend on capital investments, they need to make a profit next year to make it pay for itself. So I'm curious what you think: will these moratoriums help implement data center projects on terms that are beneficial to communities, or will they actually only slow down the process and create economic problems associated with the "bubble"? Are you following this? looking at you. I don't know. This is beyond my competence.

Yes, I'm definitely not a financial expert. I, of course, look at the. political aspect: will they really slow down the process a lot, or do you think I think so. I mean, people in the AI industry are very scared, especially because of the technical aspects. When I talked to data center computing experts after my trip, I said, "Hey guys, everyone really hates data centers." Did you know about this? And they said, "Yeah, but we'll just move them all to Texas." This was one of the most popular answers I heard. And then, after the first pauses from Greg Abbott or ERCOT, people said, "Oh, it's just showmanship."

He just has to do it as a symbol. It seems like a political gesture, but it is not actually going to restrict the activities of data centers. I've heard the same thing from industry representatives many times regarding Texas and, to some extent, Pennsylvania. I think these election results have really undermined the confidence of the industry, and I think it's entirely justified, as the political risks have become much more serious. If you read Anthropic's-1 report, you'll see that the number one risk is "we won't build the best model " or something like that. Number two-we won't get our computing power. And why can't we get them? There are a million reasons, but one of them is political risks. And I believe that

data center developers have systematically greatly underestimated political risks. Um, given my conversations with these people over the last year about the opposition to data centers, well, you know, there have been efforts to diversify. Of course, there are also Elon's space data centers. Um, of course, there are also international computing powers, which are more important now than before. Um, and hyperscalers, artificial intelligence labs, etc. Everyone is already investing in these areas, but there is no guarantee that international computing power will be more politically loyal It was interesting that after publishing my data center report, I got a lot of feedback from people in the UK,

Australia, etc., saying, "Usually, American political trends are exported to us after about six months, and we're just as concerned about the same dynamics," or in some cases, "We're in Australia and we're going to struggle with all this computing power here." So I feel like this really poses a real threat to infrastructure development. Yes. Greetings. I am David. I'm curious, continuing this topic. Interestingly, the focus has been more on moratoriums, which are at least nominally temporary, whereas in other industries that have recently become popular, such as sports betting, no one is saying, "Oh, we need a moratorium on gambling."

They say, "We need to ban it completely." Why do you think the resistance to data centers has taken the form of temporary pauses Yes, I probably should have answered that part of the question more clearly. I think for some politicians who introduce moratoriums, it's about more favorable terms of deals. In particular, if we compare agreements between communities and data center developers in 2023, 2024, and five or six years ago, today's agreements are much more beneficial for communities than they were before. Previously, small towns were forced to subsidize Meta, you know, for a data center that was supposed to give them a gigawatt of computing power.

I find this absurd. They were using the Amazon HQ2 strategy of saying, "Hey, Port Washington, you're under a non-disclosure agreement, so you can't talk to anyone about this." But unless you give us huge tax breaks to build a 750-acre data center, which is much larger than the typical internet centers, which is what people are worried about, we will raise this issue with your protege again. And these politicians, I spoke to some local officials, they were really scared. They said they wanted tax revenue, but they don't know if that's true or how much money it would bring to that company. So they simply agreed to these, frankly, terrible conditions. And so, in my opinion, there is a huge imbalance of power between

communities and data center developers. I think when I spoke to Katie Gokul on the phone about two weeks after her moratorium, which was the first statewide moratorium in the country, she told me that she didn't want it to be a permanent ban. I am a governor who supports business. I've always supported business, but it's clear to me when I talk to mayors across New York State that they don't know what the right conditions are. There's an inherent imbalance of power between the OpenAI lawyers who define the ideal conditions and, say, a part-time mayor who makes $3,000 a year and, you know, works as a tennis coach.

These are the people who negotiate, and I personally would like to see standards for making deals set at the state level. Because I believe that the states of New York, Michigan, or Texas can find faster solutions for all parties, so that they don't have to agree on all the standards, while providing communities with tax revenues and benefits that remain financially viable for AI companies that desperately need computing power. So, I think a lot of the moratoriums, there's a hope that when the political passions subside, we can move to a world where communities get standardized and profitable deals, which will, to some extent, calm people's concerns about democracy and being "cheated" because they know, "My governor, like Gina Inojoza, will fight for me and

demand the maximum to ensure justice." And if people start to believe this, they may not like the data center, but at least they will perceive it more calmly First of all, for those who haven't watched or listened to Jasmine Cesar Klein's podcast, I highly recommend you do so. I became your fan and follower when I listened to it a few months ago. Secondly, this is a question for Andy. Andy, I know you're wearing a Stanford hat here, not an Anthropic hat, but last year in your class we looked at the Anthropic Constitution, and I, for one, was a bit skeptical about a self-regulating constitution that was created and then implemented by the same company.

I'm curious, as you've gotten closer to this over the last year, how your thinking has changed about these kinds of self-regulating contracts and constitutions from some Frontier Labs, and what should exist that doesn't already exist, from that perspective. Excellent question. I am still at the beginning of my journey. I will say less than let me just summarize what I said last year and wrote in my Substack. You can see it. My concern reminded me of the work I did in social media before AI: if a technology is so important to society, it becomes problematic if a company decides things that matter to society about how that technology works. So we worked a lot on this at Meta.

We created something like the Supervisory Board, which was, well, quite interesting. She has legal authority over the company to make decisions about whether to leave content that has been removed or to remove content that has been posted. The idea was that it was a kind of experiment on how you can start involving society in your governance And last year I noticed that Claude's constitution or OpenAI's model specifications or Gemini's principles and Grok's "vibes" had the same problem, right?

They seemed to be establishing a set of rules, but it was all written by the company alone. And I think ultimately we'll want to have a system where other people are included in the decision-making process. One possibility is something I still don't know. One possibility is that this will be done through normal government policy channels. And there were discussions. I am not familiar with these discussions, so I am just reflecting as an outside observer. Discussions on proposals for self-regulation of the industry as a whole. For example, Demis Hassabis published an article with such a proposal. This can be seen as a connection to this question in the sense that perhaps there is some

authority outside of the labs that helps determine what these rules should be about the behavior of the model. Maybe it's something more democratic, where users or voters are directly involved in creating these rules, like, say, in the case of community notes or something like that, or maybe it's another option that we haven't seen yet. I don't think we know where this is going, and one of the things that has happened with the board of supervisors is that their scope has been somewhat limited. It was to some extent, I don't fully understand how it works today, but it was enshrined in EU law, but not in the US. There are no other industry participants there, and it really, I would say, hasn't reached the necessary scale, so to speak.

I think that in terms of models and these rules, we need to find a way to ensure that this happens, and I do n't know yet what that will look like. Okay, we're a little off schedule, so thank you to our wonderful speakers Jasmine and Thank you, Jasmine.

More Entertainment Transcript