Opposition to AI data centers has become the hot 2026 campaign issue. But many data centers are already being built, especially in places with less capacity to object. Will this moment be the start of feasible AI regulation or just empower some to shift development out of their area? Vasundhara Kaul finds that data centers tend to be built in minority communities and places hurting for economic development. Daniel Schiff joins us to discuss the broader implications of the backlash for AI governance. They both say better regulation is possible but does run up against a traditional political constraint: that the powerful tend to offload harms to those without the power to object.
Guest: Vasundhara Kaul, Purdue; Daniel Schiff, Purdue
Study: “Digital Footprints”
Transcript
Matt Grossmann: Data center expansion and the AI backlash this week on the science of politics. For the Niskanen Center, I’m Matt Grossmann. Opposition to AI data centers has become the hot 2026 campaign issue, but many data centers are already being built, especially in places with less capacity to object. Is the backlash a case of typically hyped not in my backyard objections or now part of broader AI worries? And will this moment produce feasible regulation or just empower those with the power to shift development out of their area? This week I talked to Vasu Kaul of Purdue University about her co-authored paper, Digital Footprints.
She finds that data centers are more often built in minority communities and places hurting for economic development. Her co-author, Daniel Schiff, also of Purdue, joins us to discuss the broader implications of the backlash for AI governance. They both say better regulation is possible, but does run up against a traditional political constraint, that the powerful tend to offload harms to those without power. We also talk about parallels to housing development politics, environmental debates, and prior technology regulation. I think you’ll enjoy our conversation. So tell us about the findings and takeaways from your paper Digital Footprints.
Vasundhara Kaul: So the big question that we sort of ask in the paper is really whether the physical infrastructure that’s underlying the digital economy, which are data centers, are they distributed randomly across the country or whether they’re socially and politically patterned, right? And we find some really interesting results. So the first thing that we found is that data centers tend to be heavily concentrated geographically, right? So we have the top five states with the highest data center counts were Virginia, California, Oregon, Texas, and Washington, right? We also find that counties with larger non-white populations have significantly higher expected numbers of data centers.
And this is after, even though we’ve sort of accounted for a range of other characteristics such as population and land availability, preexisting environmental conditions, even the state tax policy. So there is definitely a meaningful association there. We’d look into the economic story as well, but our findings are a little more complicated there where we know that the poverty and median household income that we measured behave in ways with what we would expect with the economic disadvantage argument where counties with higher levels of poverty or where there’s a high number of proportion of people have lower levels of household income, those counties tend to have higher number of data centers.
However, with the unemployment count that we looked at, we found that the proportion of unemployed residents that we measured unemployment as the proportion of unemployed residents in a county, that goes in the opposite direction to what we had expected. So a higher number of data center was associated with a higher proportion of unemployed residents, and we are hoping to explore that a little bit more as the paper and the project goes forward, right?
So our larger takeaway from this entire papers was really that it’s not that we’ve identified a single cause of where data centers go, but that we’re trying to make this argument that the geography of data centers is actually socially patterned, right? And that is why the siting of data centers is actually a governance question also, because we are asking who receives the benefits of this digital infrastructure and who’s bearing the costs of these infrastructure decisions and who has a say in these decisions.
Matt Grossmann: And Vasu, you went into this with a theory of about the local benefits being modest compared to the local detriments and maybe the potential benefits were more widely spread. So talk about how you approached this theoretically and if it’s just best perceived as a traditional siting decision like all others, or if there was something specific about data centers.
Vasundhara Kaul: What we were trying to get at with this paper was that we do see some patterns being replicated here in terms of when we look at previous literature of the siting of coal plants and landfills, right? And so we do see that in terms of the literature shows that these past infrastructure projects tend to be concentrated in non-white communities, right? And so we see that kind of replicated over here with data centers, but we also note that there are unique characteristics to data centers as well in terms of the state level taxation policy playing a really important part here in terms of determining, okay, because we found that states that do have stronger incentives tend to have a higher number of data centers located with them.
Matt Grossmann: It is sort of a local siting decision, but as you mentioned, there were state level tax incentives and now we’re seeing calls for state moratoria. So this has played out at a state level as well. Is this something where you see a real difference between the kind of local interests and the state interests or is this something where you’re just going to see people turning against or for, whether it’s by them or not?
Daniel Schiff: So I think two things. One, is this a story of classical NIMBYism? Is it local politics? Or as you said, Matt, is this a state issue? I think this goes all the way to international, right? A lot of the arguments being brought up in the AI space are data centers help secure American strategic power over China. Now whether this is salient is a different question, but has this sort of mix of scales, which is quite interesting, maybe even more than energy decisions, certainly more than housing I would say. And then I think the second thing that you alluded to in your last answers, Vasu, is that we think our conceit for this paper is this is not just an economic decision in terms of the siting of data centers, it’s also a political question. So that’s what we were trying to investigate.
Vasundhara Kaul: And I’m glad that you actually brought up the conversation around state moratoria because I think that what data centers do is show us that… Expose a classic problem of multi-level governance since many of the consequences of data centers are not actually bounded by the municipality where the facility’s actually being constructed, right? So a city, a local government will make the zoning decision, but the electricity that’s being supplied through a much larger grid and the water systems can also have cross jurisdictional boundaries. So all of these questions about where is the electricity going to come from? Do we need new generation and transmission infrastructure? Who is going to pay for that? And where are the regulations? All of those are decided at the state level, whereas the tax incentives and also the tax incentives like you mentioned, Matt, are also often created by the state governments, right?
We try to make this sort of argument that there’s a mismatch between the level at which some sort of siting decisions are being taken place and also the scale at which the consequences are being experienced. So a municipality might approve the facility, but the conversation or the effects on the electricity infrastructure or the water resources could be more regional or statewide, right? And that’s why I think we see also this conversation around state level moratorium because you might assume that data centers are more of a local issue because the zoning and the permits are happening there, but there are also all of these factors which are operating at the state level. And then as Daniel said, that there are also international considerations that often come in.
Matt Grossmann: So we just held a panel on data centers in Michigan and we had folks that had been working on economics and engineering of these for a decade or so. And basically they were like, “Why do you all care about this so much now? I’ve been doing this for a while and it was pretty quiet.” And then we had a fun experience where one of the first questions was, “I’m not seeing any benefits I can take back to my local community.” And then the very next question was, “It sounds like you’re trying to buy us off. You must be doing something bad.” So why has this changed so much, Daniel? And how should we think about the current perception of the data center debate?
Daniel Schiff: Yeah, this is an interesting and messy change in the last year or so. And one of the areas in AI politics that I’ll say I did not totally predict, I was thinking maybe labor disruption would become more polarizing earlier. I think there are both maybe objective and subjective or perceptual factors that are going into this, though I don’t think we know the full story, but maybe a few things. One I think is scale and physical visibility. So this could be, there are more data centers. I’m seeing them come up. I’m hearing proposals come up. It could also be the subjective of there are more ads about this, negative ads or positive ads trying to promote these things. So more scale and visibility. A couple of the other things that came to attention certainly would’ve been things like electricity prices and people just being under economic strain generally, being worried about cost of living.
This may be intersecting with how energy usage is playing out during the U.S.-Israel-Iran war. I think that the, we call them policy entrepreneurs, the civil society groups who are working here have been pretty active and they may have gathered a little bit of energy. We are having some conversations with them. And then one of the most interesting pieces for me about this is how much this intersects with AI generally as a policy issue. So are people saying, “Hey, I’m worried about my job. My skill’s changing. I’m a student. I’m worried about my future.” Is this intersecting with AI as a fear mixed with data centers as the physical instantiation of that fear?
Matt Grossmann: Vasu, what you studied here is kind of the end product, the ones that get built. And so that could incorporate where they were initially wanting to go or also where they were able to go. And that kind of leads to the questioner had about how much leverage do local governments have here? Is this a case, or local communities, is this a case where because these tend to be sited in particular areas, that means that there’s leverage to extract community benefits or change designs of these to match local concerns? Or is this only certain communities are targeted and that they’re targeted because they’re an easy place to build?
Vasundhara Kaul: In short, I do think that local governments do have a lot of leverage, right? And so particularly because there are all of these questions around a project requiring land rezoning or the permits, the infrastructure commitments and other sort of discretionary approvals that are taking place at the local level, right? So I think that local governments do have the power to potentially negotiate over the location of the design, the kind of environmental safeguards the data center ends up establishing within the facility, as well as the broader community benefits, right? Because we are here in Indiana, I like giving the example of Indianapolis where there was a sort of large Google data center that was going to be built, but then there was a lot of community organizing around it.
And then we had the support from the local governments as well, the community organization. And so they were able to prevent the construction of that data center. So the project was never approved, right? So I do think that state governments have a lot of leverage and a lot can be done. And I think that’s what the moratoria also kind of are indicative of, where people are trying to figure out what is the best way to navigate all of these infrastructure proposals and then also not have individual local governments make these sort of decisions where one city’s approving, another city is not, and having a more statewide kind of direction to how they’re going to approach data center proposals.
Daniel Schiff: Just to put a scary lining on this, there was actually some local political violence here in Indianapolis. There was a shooting at a council member’s house and a note left that said no data centers. So this is riling up people.
Matt Grossmann: So Daniel mentioned that this might be part of a broader AI backlash, but at least the polling as of May or June from Gallup where they asked people what their concerns were about data centers tended to be traditional development concerns, especially environmental impacts and then quality of life kinds of concerns. Vasu, some of these are things that might come from any development. Others might be more specific to data centers, but we also had a sort of back and forth where water concerns come up very frequently and yet that might be partially a product of misinformation. I know there was a incident with the author of Empire of AI that over…
… Incident with the author of Empire of AI that overstated dramatically the potential water impact of a Google data center that led to some pushback. So there are both real and real, but maybe imagined scope concerns. How should we assess these baseline concerns and how many are about data centers specifically versus just generic concerns that people have about development?
Vasundhara Kaul: So I think the answer is actually what we are seeing is something which is quite layered. So I think part of it is an anti-AI backlash that we are seeing where people are concerned about how, and I think Daniel mentioned that earlier on about they’re concerned about what’s going to happen with their jobs and how big tech companies are changing their communities. So I do think that there is part of that, but I don’t think that that’s the entire… That’s not the entire dynamic that’s taking place over here. So actually, because a lot of the conversations that are taking place are more around the politics of infrastructure. So there’s questions around land use, there’s questions around property use, like you, yourself, mentioned, water resources. And these questions around the politics of infrastructure and politics of sightings predate AI.
So I think that what’s important here is to really ask the correct questions. So if we take the water use example that you mentioned, I think we all know that data centers use a substantial amount of water. I think where it becomes a tricky area is to take that question and then move to whether data centers are creating the same water problem, which they are not, because water requirements vary enormously across different projects depending on the cooling technology the site is using, the facility design, the kind of electricity that’s powering the facility.
So I think it’s also about asking the right questions and we see a lot of community organizations trying to do that education. So they have come out with data center toolkits where they’re trying to communicate all of this information to local community members so that they can not only participate in the process, but also ask the correct kind of questions.
Daniel Schiff: This is hard for the public to get straight. It’s hard for the experts to get straight. I think this is part of the layering you’re referring to, Vasu, just the information sharing of this all. So are we saying, “Hey, some data centers don’t use a lot of water or they recycle it, but it depends on the region. Well, actually, the water is being used for electricity generation.” Well, this isn’t much compared to livestock water usage, but there are some other industrial projects that we had here in Indiana. A lot of your water is going to be diverted to another county.
And even on the research side, we have industry coming out with, “Hey, an average chat prompt only requires a teaspoon of water, but wait, is that counting for the training costs of the next model? And what if you’re using many agent calls in sequence?” So whether you’re trying to be objective about this or you’re trying to raise concern or lower concern, it’s hard to know what to ask at all or what would be the right metrics.
So when someone raises something like water is under stress, I think it’s a natural public response to say, “Oh, wait a second, this is really important to us. Slow down and reassure us.”
Matt Grossmann: Well, that’s a little bit optimistic, but I guess I wonder, and it suggests that there could be answerable concerns. So if there are noise concerns and there is technological adaptations, if there is water concerns and there’s technological adaptations, if there is power concerns and there are ways to demonstrate net positives for the grid, then all of that suggests that these are answerable concerns.
But the direction of public opinion doesn’t necessarily suggest that. It might just suggest that there is going to be opposition. It’s going to be expressed in whatever way it is. Any hints which way we’re headed?
Daniel Schiff: And I think you need to build trust there, as well. So when you have all this mixed messaging and uncertainty, then I think often the public is going to default to the more concerned position, even if there’s a technological solution that’s possible. And when you don’t trust that an industry actor is, are they really doing this in a sustainable way and now, they’re making extra commitments, the legislators are promising stuff. Even when these are positive commitments that should make you feel better, they don’t.
To answer your question more directly, the pulse is not super positive on these. There’s bipartisan opposition. It’ll be a little bit stronger or moderately stronger amongst Democrats than Republicans, but you don’t see too often this level of bipartisanship for issues. And so that’s a bad sign if you’re trying to promote data centers. So there’s a wiggle room on what policy tools or how strict we want to be. And some of this appears to shift with things like political administrations when my guy’s in power and it’s not too bad and we’re trying to test some of this stuff looking at partisan cues and so on.
But overall pulse, lots of concerns, lots of bipartisanship pushback, thus making this into a potential wedge issue in the upcoming elections.
Matt Grossmann: Yeah. So where are we on polarization here? You mentioned that there’s bipartisan opposition, that it wasn’t traditionally that partisan of an issue. On the other hand, we’re in the middle of an election campaign where lots of Democrats, including in major elections, are running on opposition to data centers. Many politicians that just a few years ago supported tax incentives for data centers are trying to run away from it. But President Trump has hitched his wagon to it in a way that might, through normal political science, might make us expect that this will become a more partisan issue. What do you think?
Daniel Schiff: That’s a really tough question. I think the role at the top of the party has presented this challenged or understanding of policy processes for the reasons that you stated, Matt. Do we see a sustained effect where a lot of the party gets behind topic X or topic Y, including support for data centers or we say, “Hey, we’re right in the throes of an election. We have lots of local pushback. I’m not paying too much of a penalty for opposing these things. Frankly, I can provide the incentives and then voice opposition so I can move around a little bit on those issues.”
Maybe a hypothesis here would be that local actors have a little bit more flexibility and they’re worried about their own prospects while the President’s approval ratings are sort of modest. General notion here would be that there’s going to remain bipartisan opposition. There’s everything from the populous right to rural populations to classic democratic strongholds. So I would suspect that this will have more leverage, unless there’s some really sustained ongoing messaging coming from across the Republican Party to shift things in another direction. And there’s a lot of baggage to address there.
Matt Grossmann: So give us some long-term perspective, long-term for this short-term debate, but where do data center public opinion fit with general perceptions of artificial intelligence in a broader set of public opinion? Is this something that people have just started paying attention to or are there long-term concerns here, and was any of this polarized before this year?
Daniel Schiff: Yeah. Data centers weren’t typically one of the questions that was raised in polls in most previous years. You can think of AI policy as going back roughly 10 years, roughly 2016, not a lot of polling back then. Maybe labor disruption, maybe autonomous weapons, do you support AI in healthcare? And part of the AI opinion story, because it’s also been quite messy and mixed and evolving. So this is to say we have people who use AI a lot, but they don’t like it. We have people that are both very optimistic about AI and very pessimistic, or they use it and they’re very optimistic, but they still really want regulation. So it is a really mixed issue area in this way. I think the public is figuring out their views.
But there are some, I think, pretty stable trends when you look across the different measures and versions. I think in the US, for instance, trust is going down, concerns are going up. You’ll get many more people will say, “I’m more concerned than excited.” And we’ll say, “I’m more excited than concerned.” Labor disruption, worries about job loss, that’s been stable, but increasing. People have been worried about that for a long time, even though they tend to be more worried about the economy than their own job in particular.
So general pulse there, I would say, is lower trust, more concerns. People still like it in some ways, and excited. They may not want to shut down AI development, but they’re unhappy with the lack of regulation. And that is, again, bipartisan.
Matt Grossmann: So yeah, you found that you can have opinions on governance that are separate from your just pro or con. So walk through those and tell me what is a just broadly supported set of policy ideas surrounding AI, and then where does the more controversial or polarized opinions come in?
Daniel Schiff: Yeah. Okay. So some broadly shared views on AI. So whichever major party you’re asking, you’ll get some net support for you. So I still like AI, I still want to develop it. I think it’s cool for some things. Definitely AI for healthcare, I support that. Less AI for autonomous weapons. You will get at the same time strong consensus that we want this to be regulated or carefully managed. Sometimes we’ll use different terms here if people are not happy with the term regulation. That again, that’ll be bipartisan. We’ve seen this amongst the legislators. I’ve surveyed state legislators for instance, and we asked them about the balance between innovation and human rights or should we focus on being safe and ethical or out innovating China? And there’s broad interest in “being responsible” and earning trust and doing careful regulation.
So where we are right now is pretty far away from probably the floor of what people want, in terms of having more well managed and safer and better regulated AI. When we get to that level, if we get to that level, then you might see more polarization. So for something like data centers, do I want to stop them and ban them or do I want to just ensure that there are sufficient community benefits and environmental controls? So you might see some divergence over policy tools. But yeah, generally the public is lukewarm, but mixed about AI, but not happy with the state of regulation.
Matt Grossmann: Vasu, what do you think the lessons from the data center backlash are for the broader AI policy debate? On the one hand, your findings come from an era in which most policy was pretty positive, and yet, the results show that the benefits and costs didn’t go equally across the population, but we’re now in a different time in terms of both public opinion and policy proposals. So what do you think the lessons going forward are?
Vasundhara Kaul: Yeah, so I think that one of the key motivators or findings in terms of policy implications of our study and the subsequent studies that we are planning is that AI governance can’t really stop at algorithms and models. So there is this physical infrastructure that tends to get lost in the conversation when you’re thinking about the cloud, because we think of it as this placeless entity, but there’s all of this physical infrastructure that is supporting the cloud. And there are decisions about natural resources, electricity, taxation that go into regulating this physical infrastructure. So I think that’s what we are trying to demonstrate, that why it’s important to study physical infrastructure and why it’s a governance issue.
I think that we also show that it’s important to think about the cost and benefits and how there’s this dissonance between, like you said, there’s these localized costs, but the benefits are more dispersed and how certain communities are disproportionately bearing that cost. So I think that policymakers really need to think about the distribution, and not only just think that overall, what is the aggregate cost benefit calculation that’s taking place.
And lastly, the main aim of our team, and I think that something that is extremely important in terms of policy is that community plays a major role in all of this. And that’s why we are particularly trying to study community voice and what role that is playing in these governance questions around physical infrastructure, and community participation has a large scale implication on how these projects are designed, where they’re eventually located, and the level of acceptance or resistance that takes place against them.
So these are the three big takeaways that we are trying to emphasize in our work.
… takeaways that we are trying to emphasize in our work overall.
Daniel Schiff: So you mentioned, Vasu, we think about AI governance beyond the model layer. This is something that might seem straightforward to many of us, but so much of the AI governance conversation is still dominated by technical evaluation of algorithms. I saw this as recently as a couple weeks ago in DC, and many of us have been saying for the last decade, “Hey, this is a sociotechnical problem. It’s a collective problem. There are different types of risks.” And still a lot of what’s being emphasized is deep fake detection or red teaming safety valuations. And this one is even more distributed and complex in terms of thinking about that cost. So totally agree with that point. And a second thing you raised to me is… Maybe I think one of these lessons is industry to some degree, policymakers to some degree may not have paid enough attention to public trust and participation.
There’s a lot of gesturing at this. Tons of documents, government, nonprofit, private sector. “We care about public participation. We care about public voice. We care about diverse participation,” over and over again. There’s one media article that says this is the most common recommendation in AI policy. And yet a lot of the discourse in my mind has been about, “We want public trust and then that trust will bring adoption.” And not as much about, “How do we actually earn public trust? How do we not just have trustworthy AI, but trustworthy AI governance?” And I don’t think we have either of those. So to the extent that people are just pushing forward with AI or data centers or whatever and thinking, “Yeah, the public, they’re going to get on board. We like this. Can’t we see this as good?” Not paying attention to that, I think, is creating a debt that these actors are now seeing.
Vasundhara Kaul: Just one last point that I wanted to make, sorry. But it’s just that we also tend to think of all of this in terms of individualized trust in terms of individualized benefits. So even if you look at something adjacent to it, like privacy, so we’re thinking about, “Okay, what steps can the individual take to protect the data privacy?” Or, “What steps can we take to encourage individual trust?” But I think what these sort of [inaudible 00:32:25] data centers really helps us refocus on is that these are broader collective issues that need a more collective response and a collective framework when we are studying them as opposed to only looking at what individual preferences are.
Matt Grossmann: So you both called for more community voice and I think we have a lot more community voice now. But Vasu, as you’ve mentioned, the patterns you find follow all kinds of normal patterns in environmental justice research. They also follow patterns in things like affordable housing that many people actually attribute to voice. That is, you get these patterns because when you empower local communities, what that means is that the people who are able to show up to the meetings and the people who are lawyers and environmental scientists are able to drive the discussion in their communities, but other communities are likely to be the place that you get the things that those communities don’t want. So is there a downside?
Vasundhara Kaul: So I think that is a tricky question, I think, because like you said, that there are certain communities where you would find, say… Create a proportion of environmental scientists or people who have the time and the resources to get involved in community organizing. So I think that is one of the limitations where we are just saying that all the onus is on the community to organize and resist data centers. And I don’t think that that’s the argument that we are trying to make here. I think what we’re trying to say is that it’s important that the onus lies on local and state officials to account for community voice. So the onus is not only just on the individuals to go and dedicate their time to making these arguments for the protection of their community. I think the onus is on the policymakers or the officials to actually make the effort to hear community voice and then act in adherence to that.
I know it’s a very idealistic scenario, but I think that’s what we are trying to say by emphasizing community voice. And yes, in certain situations that community voice might be absent. And I think that’s where the important role of local community organizations come in where they’re stepping in to create that awareness and do the advocacy on part of people. And I think part of our analysis actually… We’ve not been able to demonstrate that as currently in our analysis, we’re only looking at data center count. And so these are already data centers that have been constructed. And so it’s possible that these data center were constructed in these areas is because there wasn’t sufficient opposition for them, not because of any other factors, but because there wasn’t sufficient opposition to them. So that’s why certain areas do have a larger data center count. And we will try to address that in future iterations of this study.
Matt Grossmann: The way the energy siting policy has advanced is through a permitting reform at the moment to actually reduce the number of processes that people have to go through for energy siting in the housing debate. Again, the solution has been to go up a level of government to force local governments to approve more things or incentivize them to. So in both of those cases, people have actually been moving toward, I guess, solutions to community voice. Does that say anything about this debate?
Daniel Schiff: I’ll try not to punt on this question, Matt. I don’t think I have a good solution to this paradox of participation. I think this is what I would expect is to see NIMBYism play out in ways that perpetuate economic injustice as it were. But I want to propose a broader version of what community engagement looks like here. So one is, yeah, maybe we have a societal conversation and we decide collectively, “These things are good. We want more of them, but we need to reduce the regulatory barriers. That’s best done at the state.” Or we say, “Hey, look, these things could be good, but we’re worried about water or land usage or local benefits and tax abatements,” and therefore the companies are going to come out with their own statements promising to be more responsible and the local policymakers [inaudible 00:37:33]. So we established these channels, these debates, and this just makes everyone do better policy. I think this is a really viable and important piece of the voice here.
Matt Grossmann: So I also want to get to a paper that you recently wrote, Daniel, on issue polarization in some of these debates because you outlined some mechanisms by which issues are polarized. And so maybe it would serve to go through those and say how many apply to this issue.
Daniel Schiff: Hey, sure. So this is a pretty recent paper, but it’s looking at early development of bipartisanship and then maybe emerging partisanship in AI policy broadly. And you can think about this issue historically as pretty bipartisan in the US. Both major parties, they want to innovate and they want to beat China. And that’s more or less the shared messaging for a while. They want to be responsible, innovate, beat China.
To different degrees, I think one of the big Senate and House bills, one mentions China 100 times, the other one mentioned it 500 times. So a little bit of some people like environmental issues a bit more. But you kind of have a general story. But then we start to track and say, “Hey, when is this going to become polarized? Will it? On what basis?” And the framework that we come up with here lists four different dimensions where we might see polarization.
So first we have problem definition or problem framing. So someone comes along often strategically and they say, “Hey, AI. AI has taken your jobs. It’s bad for your jobs.” Someone else says, “This is a racial justice issue. AI is discriminating.” Someone else says, “We need the military advantage over our competitors.” So there’s a bunch of different ways to talk about this. Or, “Hey, AI is fine, but we want it to be transparent. Everybody likes transparency. This is a shared value.” So I can kind of pick issue frames that are more polarizing or less polarizing. And I think we see different actors do that.
Second dimension is the policy tools or policy solutions. So we can sort of agree, “Hey, we like AI. We want to do it responsibly. What are we going to do?” And we might have a very minimalistic, “Well, you at least need to list your AI use case on a government website. We have basic transparency. We’re going to tell you that there’s a chatbot that you’re interacting with.” Or we say, “You have to have a mandatory audit and bias assessment and we’re going to block you from building your data centers so we can have stronger or weaker policy tools.”
And I think in a number of cases we’re seeing there’s agreement on the easiest or thinnest policy tools, but when we want to push harder, then you see the Democrats and Republicans splinter, and that’s something you might see for data centers. Third and fourth, I’ll keep these shorter. Third is the actors or stakeholders who people listen to. So who becomes powerful? Who becomes organized? A few years ago, you didn’t have these big frontier AI companies as heavily engaged in lobbying. They didn’t have expertise. They weren’t doing state level lobbying. Now you see them mature, and you can see this in different sectors. You can see bail bondsmen arguing for or against algorithmic audits. You can see this in the criminal justice sector. You can see it in education sector. So which actors take hold.
And then the last piece here is about what we call the policy system or subsystem. And this is the idea that AI is such a general purpose technology, it touches every sector that again, you could say, “AI is an economic innovation issue, it’s about development.” Or you could say, “It’s about immigration and STEM and who’s coming into the country and we want higher or low skill workers.” Or you can say, “It’s a military issue.” Or, “It’s a justice issue and it’s taking away the meaning from our jobs.”
So you can put this in a bunch of different sectors and maybe depending on how you’re able to frame AI as a broader issue, you can mobilize different constituencies. These are sort of four different dynamics that we think are in play that can help explain movement towards polarization.
Matt Grossmann: So many of those sound like our traditional kind of elite led public opinion models of polarization, but doesn’t the data center backlash suggest that there are still bottom-up processes where the public comes first and elites actually follow?
Daniel Schiff: I think there is definitely a room for top-down and bottom-up. And one of the questions in AI is who is going to lead here? You might think classically this is a highly complex technology and even more than normal, we have the black box problem, so even the experts don’t understand what AI systems are doing. So maybe we’re really going to listen to the elites and experts here. And we do see some of that when we poll state legislators, “Which actors do you think are most important?” But people, they will express at least that the public is important. They have voiced this in many ways. To that point, I think this raises the earlier concern that some policymakers, some actors in industry have gestured that they care about bottom-up opinion and this is going to matter. Trust is going to matter. If doctors are going to adopt this, if teachers are going to adopt this, if people are going to drive self-driving cars, but then not really taking public opinion seriously and not earning trust.
So I think this is under indexing on how much the public matters. I suppose I would say with respect to our dimensions in our study here, I would imagine some of them to be possibly led by the public. So probably the public isn’t coming up with the policy tools that get on the agenda, but the public is shifting what kinds of problem frames become most salient and they’re responding to the policy tools that policymakers are proposing.
Matt Grossmann: So as you all have experienced in the last couple months, this issue has dramatically risen in salience alongside broader concerns about AI. Some media tracking has these mentions of these issues quintupling or so in very recent periods. So how is that likely to shift these debates?
Daniel Schiff: Are you thinking here of the sort of AI safety hacking catastrophic risk conversation or something else, Matt?
Matt Grossmann: Yeah, it seems like what happened is that we had a kind of a slow build of the data center backlash getting more and more attention. And then suddenly we had a much more sudden conversation about existential risk in AI and these things are going together in some people’s minds, but I don’t know that that means that it will change the data center debate. But typically when something becomes much more salient, you see a different kind of debate.
Daniel Schiff: So I don’t know that… Let’s say there’s been a merged policy image of the AI safety concerns and the data center concerns. My suspicion is there’s some latent connections about overall AI attitudes that are feeding into data center attitudes and vice versa. And that’s something that we are hoping to measure and maybe some people have started.
The AI safety issue is interesting. So we have kind of a whistleblower who gets a lot of attention. For reasons I’m not totally clear, maybe this is just because it’s following the Hugging Face incident and that was pretty visible as an incident. But this has been the concern of AI safety people since the early 2000s, since I started work here. And the frontier companies, this is what they’re all about. The mission of these companies is to build artificial general intelligence, super intelligence. Many of the founding figures are deeply in this camp worried about catastrophic risk. So this is their DNA. I suspect this is also part of strategic framing, the AI-
DNA. Yeah, I suspect this is also part of strategic framing. The AI safety community has gotten more skilled in lobbying and advocacy and persuasion has funded education and communication efforts. So I think they’re certainly, again, thinking about what types of policy subsystems they can gain traction in. Connecting this to data centers still strikes me as tricky.
But yeah, I mean, we have a moment where there is bipartisan concern. There’s actually quite high levels of concern in some polls about what we call catastrophic risks from AI, AI takeover scenarios. We have a letter that came out from 20 presidents and prime ministers, and we’ve had all sorts of statements. Stephen Hawking, Bill Gates, Elon Musk go way back, pause letters about AI. So maybe we’re starting to see enough crystallization of public concern since the launch of ChatGPT that we’re getting to a fever pitch of like, “Hey, let’s regulate stuff now.” And you’re throwing data centers at us too.
Matt Grossmann: The other trend has just been the rising economic influence of the data center build out. And some issues that don’t necessarily come up in the local siting, but have increasingly come up from economists are things like crowding out of other debt or having a market that is for both debt and stocks that is increasingly concentrated in the success of the AI build out. With lots of discussion about … Originally a lot of discussion about a bubble, then discussion about potential crashes.
How should we view that when both of you started with the idea, these are going to get built somewhere, where do they get built? But we are increasingly having a conversation about just how much of the economy is tied up in the anticipated large scale build out.
Daniel Schiff: I’ll give a tiny little take here. And Vasu, you may know more. I’m not a macroeconomist.
I will say AI has this sort of infamous history of springs and winters or summers and winters where people get really excited for 10 years and invest a lot of money and then they don’t make enough progress and things crash. There are reasons to think the current wave is, I think, more sustained importance that we’re already getting tools that are capable of shaping the economy.
But yeah, I mean we go through cycles in the economy, crashes and bubbles are scary. We’ve lived through several of these. The older you are, the more of these you’ve lived through. So there’s a lot of eggs in this basket here, economically, structurally, socially. So if and when some kind of bubble or crash happens, that’s a lot of people who will be picking up the pieces here. And I think the actors who’ve been promoting these will potentially feel the backlash that results from that.
Vasundhara Kaul: Yeah. And just to add a little bit, so I think people are also thinking about this. But from a preliminary insight, we do know that there are concerns in Michigan that data centers will come in and then there’ll be construction and there might be some temporary employment. But then very similar to some of what we’ve seen with a past infrastructure project that they will then eventually leave. And then that creates this sort of complete breakdown of the economic setup in those particular areas.
So I think people are also very concerned about that and something that they are considering when they’re thinking about, “Hey, do we want a data center in our community or not?” And I think that is something that should be an ongoing and very important conversation.
Because an important point that I would like to clarify is that we are not trying to say that these are just going to be built. So where are they going to be built? I think our main aim of our research is more to say that let’s look at the sociopolitical consequences of these infrastructures being built because these are not just technical requirements at everything. So people are not just looking at like, “Hey, the water is there, electricity is there, so let’s just build that.” People are looking at a lot of other factors and these have implications for inequality.
Daniel Schiff: Maybe another piece that makes this messy is it feels that this round of investment involves more uncertainty. So if I think about investment in energy, hey, we know our country’s energy supply of our regions, we’re getting more people are moving into the state. I can plan for this. I know renewables are coming and there’s some tax changes sometimes that change my planning.
But with AI, do we need to build 10 times as many data centers? Are we going to come up with faster methods in six months, like the deep seek moments where say, oh, we don’t need 80% of this capacity. And is this going to radically transform the economy and everybody’s going to be using agents all the time? Or are we going to go on some other direction? So I think the economic investment uncertainty is also tied up with the social and labor precarity.
When I’m talking to students at a panel last week, they’re worried. They’re worried about the future, they’re worried about their jobs. All of this stuff feels up in the air and all of it feels entangled. And I think maybe that’s what is creating a little bit more anxiety at the moment for this particular possible bubble.
Matt Grossmann: So political science has been kind of a bit player in the AI revolution discussion for the most part, but increasing. What are you seeing in terms of the role that scholars of politics can play? What’s the sort of state of the field and how much of a boom is there going to be in research surrounding this?
Daniel Schiff: I’ll give a little historical take and then I’d be curious how you’ve experienced this Vasu.
I would say early on as someone who started looking at this in the late 2000s, there’s not a ton of interest in AI policy coming out of political science. Public administration was probably first. They had this history of caring about e-government and being kind of close to the ground.
And then public policy, they start to look at it in the late 2010s a little bit and political science I think really starts paying attention post-ChatGPT, which I think is the same for many other fields. So it’s unavoidable. You have to pay attention to this because it’s affecting trade, it’s affecting dating apps, it’s affecting sports, it’s affecting everything. So a lot of scholars are working on this. Much of this I imagine is driven by funding and attention. People want to focus on exciting topics, but it is influencing just about everybody’s fields.
We do have a lot to offer here, political scientists, policy scientists. One of the things I mentioned earlier was a lot of the AI governance conversation has been about technical AI safety. This community was very squarely focused on how do we create alignment within systems? How do we create mechanistic interpretability? This is a technical problem. And I remember in some of those early days saying, “Hey folks, this is going to become policy issues. The public opinion’s going to matter. We need to think about things like framing. We need to think about organizational adoption and standards.” And people being like, “This is a computer science thing. We’re going to solve this problem.”
And now it’s become just unavoidably questions about stakeholders and organizational design and harmonization of regulations. So I don’t know, there’s a not quite infinite, but a very large number of questions where a lot of work and a lot of help is needed. So I’m glad to see more people focusing on the issues.
Vasundhara Kaul: Yes. Full disclosure, I’m a sociologist who’s trying to infiltrate political science. So yeah, I’m going to give a sociology response here. But yeah, I think I’m glad to see that so many researchers are working on these issues and it is getting the importance that it should across disciplines. And so that’s lovely to see.
But I would like to note that these are questions that social scientists have been asking for a very long time. I think AI just provides a new context. So like you said, Matt, environmental justice scholars have been asking this question around coal plants and affordable housing. And I mean, sociologists ask those questions around affordable housing. So yeah, I’m not sure whether this will lead to a whole different burst of research. But I think it does remind us that these questions that we have historically been asking about inequality, about political participation, about social determinants of these so-called technical things are still very pertinent in today’s age of AI.
Matt Grossmann: And anything we didn’t get to that you wanted to include or leave us with? Or anything you want to tout about what is next?
Daniel Schiff: I’m always curious what your takes are, Matt, from our conversation, what you disagreed with, what your feeling is at the moment.
Matt Grossmann: I’m just trying to figure out how much of the data center stuff is related to normal NIMBY politics in which I have my own kind of framework to think about that, versus is specific to AI. And I think it might be changing as a result of the high salience of this issue that you saw motivating some of my questions.
But actually, I mean, I would say I’ve been looking around for people who have been doing studies of this. We had an episode about a year ago on early AI policy trends and there really wasn’t a lot. And I look in sociology too. And I feel like it’s only now that we’re starting to see a lot of stuff where it’s like, oh, okay, well, this question that you have, there exists this research agenda to study it and it’s just in the social sciences and you haven’t gotten to that. But maybe you all have more on the ground to how this is being received now.
Daniel Schiff: Two thoughts there. One, this is a question I think about a lot, again, being kind of early days is what is different about AI policy? And in public policy fields, we can have environmental policy, education policies, we can have a subject matter area which is a little bit different from how political science orients. Is this the same or is it more general? Is it more complex? Is this other trends that are changing in terms of stakeholders or industry power? So I think there’s interesting and different stuff to tap in there.
And just to give one example for data centers, the area where I spend the second most time on this is talking about war and international security. So people saying, “Hey, we’re bombing data centers in the United Arab Emirates. What do you think about that?” Now this is not totally different from bombing an oil field maybe, but it’s a little bit different from a typical NIMBYism conversation. And that’s kind of the stretch of data centers reach data sovereignty questions.
A second AI related thing that is entangled with this really is AI as method. So I’m glad and excited that all of our disciplines, humanities scholars, whoever, are bringing in their wisdom from decades and trying to get this into action policy as I value, which it’s always a fight, but we need to be doing this. But now we’re also getting a ton of research and a ton of extra research and a ton of AI generated research. And I worry about things like effective knowledge synthesis and getting effective information out there and shaping policy wisely. So we’re advancing solutions and knowledge. And we’re also just creating intentionally, structurally, unintentionally, lots of chaos. So again, AI is the sort of dual use challenge here that we’re navigating.
Matt Grossmann: There’s a lot more to learn. The Science of Politics is available biweekly from the Niskanen Center, and I’m your host, Matt Grossmann.
If you like this discussion, here are the episodes you should check out next all linked on the website. Making AI Policy, Are We Falling Behind or Rushing In? How Not in My Backyard Politics Keeps Housing Costs High? Can AI Vibe Research Replace Social Science? Can Housing Policy Overcome Lousy Folk Economic Thinking? And When and Where Can Climate Policy Succeed?
Thanks to Vasu Jakkal and Daniel Schiff for joining me. Please check out Digital Footprints and then listen in next time.
