Leading AI Professor: We Must Address AI's Climate Impact Before It’s Too Late | Kate Crawford

0:00If we don't change how we're building artificial intelligence, what will be the overall cost to the planet? If we're racing towards a climate disaster, if we're racing towards a real collapse in the labor market, we might want to be asking harder questions about how we build and use these technologies. We tend to think about these technologies as an extension of the brain, but it's an extension of Earth. Generative AI will be using the same amount of energy as India by 2030. The narrative right now is that it is AI and national security or the planet. And that lets tech companies get off easy because any decision maker or world leader is going to choose national security in this moment.

0:38We're looking at eight billionaires who are having extraordinary power over the lives of eight billion people. This has to be a public conversation. This is one of the last moments where you can really start to shift what these infrastructures are going to look like. That to me is the real urgency.

1:00Professor Kate Crawford is one of the leading voices examining the environmental, social, and ethical costs of AI and emerging technologies. She's a professor at USC in Los Angeles. She advises world leaders and policymakers. And today, we're going to pull back the veil on the planetary costs of artificial intelligence. I'm Sine Bovel, and this is I've Got Questions. Professor Kate Crawford, what do you think is one of the biggest illusions about artificial intelligence? Oh, that's a good one. There are so many illusions at the moment, but I think perhaps the biggest one is this idea that AI is immaterial.

1:40It's a magic trick. It's code in the cloud. In actual fact, we're in the middle of the biggest industrial transformation in a century. In fact, it's been called the biggest infrastructure project of all time. We're seeing massive data centers sprawling across the US and across the world that are predicted to be using as much electricity as Japan by next year. So this is an enormous energy, water and resource draining infrastructure that's being built so rapidly that some of the new data centers are being put in tents because they can't build the buildings fast enough. So, I mean, I think at this point we have to start looking at AI as a massive industrial infrastructure that comes with real costs, particularly if we're not thinking ahead in terms of how to make this sustainable.

2:32And what are some of those costs? So for somebody who is not familiar with how AI is built, how these systems are run, where does the environmental toll happen? And is it worse when AI systems are being built or when we're interacting with them? Well, I mean, if we really break it down, essentially AI models run on semiconductor chips. So these chips, they're called things like the A100 or the H100. These are enormously energy hungry and they also produce a lot of heat. So in addition to needing a lot of electricity, they also need to be cooled down, often by using fresh water that's coming from our reserves.

3:11And in many ways, these giant infrastructures can have somewhere from 10 ,000 to 20 ,000 to 30 ,000 of these chips running 24-7. Now, the way that AI models are trained, they actually go through this very large process of being trained on a data set that can be as large as the web itself. I mean, these are enormous data sets, often not very well curated either. So you're sort of throwing everything in. And that training process can take a full year or in some cases even longer. That is obviously a very resource intensive period for energy and water. But then there's this second stage, which is called inference.

3:50That's when you and I are asking questions of, say, chat GPT. Every time we do that, it's depending how you count somewhere between five to ten times more energy intensive than an ordinary search. Now, if you times that by the number of queries that are happening every day, hard to get that number, but OpenAI says they're getting around 2.5 billion queries a day. That's a lot of energy and that's a lot of water. Then, of course, there's the mineral layer. To build all of these things actually requires mining everything from rare earth to lithium to cobalt. In my work, I've done a lot of field studies going to these sites, visiting the mines, seeing what happens environmentally, but also in terms of the labor.

4:38This is some really backbreaking work in these mines, often exploiting child labor, depending where you are in the world. And then, of course, we're seeing enormous pressure being built and basically being put on these mines to produce more and more and more to build all of these new AI hyperscale data centers. So at every level, if you're talking about minerals, water, energy, all of our data, all of our work over the years, these are becoming the resources that are driving this AI revolution. So in many ways, AI in its current form is competing with humans for resources of survival. Exactly right.

5:14And in this sense, you know, it's not artificial intelligence. It's massive industrial scale synthesis. It's inference engines that are drawing so much more power than most people even realize. In fact, some analysts predict that generative AI will be using the same amount of energy as India by 2030. It's like bolting a whole new industrial nation onto the grid. And of course, you know, I think if you listen to industrial leaders, they know this. You know, they're saying that right now data centers are being planned that are 10 gigawatts. You know, that's a lot when we realize that a nuclear power station, for example, is one gigawatt.

5:55So really, these AI data centers are going to be using as much power as entire cities. So that comes at a real cost, particularly because this is not all alternative energy. A lot of this is using gas, fossil fuels. Which is the big part of it. It's the choices that are being made. And so are there examples of some of the harm that's happening right now to communities in real time because of these data centers? Because like you had mentioned, we think of the cloud as this thing that isn't really real. It just floats around. But it's not exactly that. And a framing that I loved, a point you made in your book, Atlas of AI, is that we tend to think about these technologies, especially one like AI.

6:33And we hear this metaphor all the time as an extension of the brain. But it's an extension of Earth and the planet, not us. Exactly right. And you asked this really good question of, you know, what are the lived impacts for communities right now? And perhaps one of the most dramatic is playing out in South Memphis, where Elon Musk has built what he claims is the current biggest data center. He calls it Colossus. That's a whole other conversation. Part two. Yeah, part two. But, you know, Colossus requires far more energy than the grid is able to provide. So they started to use a whole of methane gas generators to power this enormous structure.

7:17That releases quite toxic nitrogen oxides and formaldehyde into the air. And this is affecting a historically black community that already has some of the highest asthma rates in the country. So we're talking about local communities really experiencing very real downsides. That's before we even talk about the cost. I mean, we're seeing the cost of electricity be passed on to consumers. In some cases, it's doubling in some regions. And if you look at a state like Virginia, right now, 25 % of the electricity in Virginia is being spent in data centers. That's going to go up to 50 % in a few years.

7:56This has real implications all the way down for people and communities, but also for our environment and also for the kind of world that we're trying to build. But if you look at some of the more recent numbers, right? So you had made a point at a talk that we were at in February about DeepSeq. So this is a Chinese AI model that came on the scene and made a lot of headlines, not just because of some of the efficiency gains, but because it showed America they're not the only game in town. But it showed that you could build an AI system with one-tenth of the computing power than some of its competing models.

8:29And then 80 % of the energy tends to happen in inference. So when we're chatting with these AI systems or when we're making images. And some of the recent gains, I mean, there was a study from Google that showed in just 12 months, they've reduced their carbon emissions and inference by about 98%. So things are somewhat trending in the right direction. But is this not the whole story? Is that enough that we're just seeing a downward pressure on emissions in a good way? Well, I mean, let's talk about that Google report. I think it's, first of all, fantastic that they're finally releasing some numbers.

9:01researchers like myself and many others have been pushing for much more granular data. So that's fantastic. And of course, Google is saying that they've increased their efficiency really quite dramatically in the last couple of years. And if you really dig into that report, there's a couple of things that are worth noting. One is that they really only look at the numbers for text generation. So if you're generating an email, for example, they don't look at images or video that are far more energy intensive. Another thing here is to think about what efficiency really means. It means that effectively you're lowering the cost of every query.

9:39Now, this is really interesting because it takes us back to a well-known paradox in the economic literature. It's called Jevons paradox. This was basically an idea that was discovered by an economist in the 19th century called William Stanley Jevons, and he was looking at coal production. He found that when we made coal production more efficient, rather than people actually using less coal, they used more. AI is like that problem on steroids, because even if you make every query more efficient, then we start to see bigger and bigger models being built with more and more hyperparameters from billions to trillions of hyperparameters.

10:20So even if you have more efficiency, the increased use at the inference layer that you just mentioned actually means that overall, more energy is actually needed. But there's something else going on too. And I just wrote a paper with two of my colleagues, Sasha Lucione and Emma Stroubel, where we really dig into Jevons Paradox. And what's interesting is it's not just a question of energy. It's a question of our behavior. It's the fact that we're getting so used to AI being built into everything that we use every day. Even if you're not thinking about it, a chatbot will suddenly appear and say, can I help you?

10:54Even if you're not intending to use Gemini, you'll be getting an AI recommendation. So it's starting to infiltrate at every level, which is changing us behaviorally. So we have to start to look at a much bigger social analysis of what AI is doing. And I think that's really been the focus of our research recently. Right. And it's an amazing point because the more efficient the technology becomes, the more we're going to use it. The more progress we see in the technology, I mean, it's the worst it will ever be. Even if we don't see much more progress, we'll figure out new ways to use it. And I guess AI being a general purpose technology, that is the path that it is on.

11:30It is going to become infrastructure. And some of the things I mean, I find funny too when I read these reports, is that they often make comparisons to a Google prompt or a Google search in 2008 or streaming Netflix. Now we're down to a few questions with an AI system is equivalent to maybe eight to 10 seconds of Netflix. But that doesn't actually make me feel any better because we're still using the wrong energy sources, because not all energy is equal on the planet. So it's still, if we zoom out, how are we actually living in our behavior and the technological infrastructure more broadly? If we don't change how we're building artificial intelligence, what will be the overall cost to the planet?

12:12What timeline are we on? Well, this is a really hard question to answer, again, because of the difficulty in getting the real picture in terms of the data out of all of the major companies. But what we do know is that the current predictions for, say, how much electricity AI is going to need are shifting from between 30 to 90 % in the next 10 years. That is a staggering figure. Obviously, if we keep on this path of using a lot of fossil fuels in the mix, then we're accelerating a really dangerous trajectory. We're already in a perilous position with climate change. The new reports, frankly, are keeping me up at night.

12:52I know this is affecting you as well. So in that sense, we just don't have time to be building these highly inefficient fossil fuel driven data centers. And there are some really good examples of making different choices. You can build AI differently. And you can think about it down the stack. Let's think first of all about where you're getting your energy from. Now, China, for example, has got something like 900 gigawatts of alternative renewable energy in their grid. They just added another 300 gigawatts of solar just last year. That means the marginal cost of your AI queries is approaching zero from an energy perspective.

13:33That's something that the US could and I think should be doing. And in fact, a lot of us, again, have been pushing from a policy perspective to go in that direction. But in fact, we're seeing the opposite. We're seeing a sort of a doubling down on fossil fuels and a de-incentivization for following a renewables trajectory. So I think this is a really urgent issue that we need to look at. And again, to think about what this is going to do in terms of the world that we're going to be living in in just 10 years time. If we go down the stack again, though, there are all sorts of other things you can do.

14:06For example, you could start to use far more tailored curated data sets instead of these huge internet scale data sets. That's less energy. You could also use something like a mix of experts approach. That's, of course, what DeepSeek used. It's part of the reason that their energy demand was much lower in terms of training that model. That means instead of spinning up a gigantic general purpose model, you're actually just using small components of a model at any one time. And also just generally speaking, people are deploying huge, large language models for customer service where you could just create a customer service spot for 100 times less energy.

14:45So these are the kinds of decisions that I think we could be making. Part of the reason that's not happening is that we're in a race. And in a race dynamic, everyone feels like bigger is better, as fast as you can, get your model out before the next group does. But the question that's not being asked is, what are we racing towards? And if we're racing towards a climate disaster, if we're racing towards a real collapse in the labor market, particularly for younger people, if we're racing towards a system where we're outsourcing our own cognition and creativity, we might want to be asking harder questions about how we build and use these technologies.

15:26Right. Where are we even going? And I think it's really important that you brought up China because the race, I mean, America has named China as its competitor in chief in this AI race. Right. And there is a lot of truth to the national security implications of this technology and the geopolitical backdrop that it's coming online in. But I think that there is a narrative that should be corrected because the narrative right now is that it is AI and national security or the planet. that it's this zero-sum game. And that lets tech companies get off easy because, of course, any decision-maker or world leader is going to choose national security in this moment.

16:09But that's not true. It is not zero-sum. There are different choices in how to build or how to power AI systems, and America's biggest competitor is making those choices. So do you think it's time to change the conversation around AI and clean energy? It's not just for a moral imperative reason. It's in the name of national competitiveness. I mean, even when you look at China, I'm sure some of it has to do with climate. But it more has to do, I'm sure, with national security. I mean, they were the largest energy importer, one of them in the world. They were very dependent on global markets and geopolitics.

16:45And they recognized that in 2010 and said, nope, we have to cut that off. So isn't, I mean, if anything, the national security argument is really strong for clean energy. So what is the U.S. not getting that China is? Honestly, I think it's ideological. I don't think we're looking at a set of rational choices around the energy mix for the United States at this moment. And that's a real shame because we happen to be at the moment historically where we have to make the most dramatic decisions around how we're going to be generating electricity going forward. So I think if you really talk to people in the industry, you know, off the record, they'll tell you the same thing.

17:25They know that clean energy and renewables is the only way forward, not just in a sense of virtue signaling, but in a sense of actually lowering the real cost. I mean, even for people who don't care about climate change, I don't understand how they can feel that way in this day and age. But they just look at the cost of running these models. If you are moving towards solar, for example, that's raining down from above. that is a much easier option for you. So we could get into the weeds. There are some really interesting numbers and facts and figures around what you can do in terms of building different infrastructures.

17:59But perhaps the biggest takeaway here is there is another way. And if AI is going to produce important benefits to society, it can't be doing so at the cost of our ability to live on this planet. I 100 % agree. Even if you want to put the ideology of climate aside for people who call it that in the name of capitalism, in the name of the bottom line, and in the name of national security, clean energy aligns. I mean, this is technically probably the first time in history where national security needs, energy needs, and the needs of companies and for the planet are all in alignment. And these are choices that are being made to not go that route.

18:38And I think the challenges too, with this zero sum, it's planet or AI national security is that it puts individuals in an impossible position. Because now there's a growing number of people who are opting out of using AI for the sake of the planet. And that can come at a steep cost, right? You're not learning the skills, so you're risking future employability. You could be having productivity gained in your own life. And then it's the same offloading of systemic responsibility onto the individual. It's the don't use your plastic straw and please carry your paper bag to work. And it's not reasonable to make people choose between employability and clean air because it's not a personal consumption issue.

19:19It's a structural design failure. And these are two very different things. I couldn't agree more. And we've been here before, of course. The whole idea of the individual carbon footprint was invented by the oil industry to try and put all of that emphasis of behavioral change onto the individual. We're seeing the same happen with AI. To be super clear, I don't think this is down to individuals to try and remove their carbon footprint from the AI mix. It's simply impossible given the fact that you can't even elect whether you're using some of these systems. They're being built into everything you use from your phone to your browser.

19:58The bigger question is, as you say, a structural one. And it really means there needs to be a lot of pressure on this industry to move towards renewables, and also to start designing AI for the long term, for its sustainability, for human flourishing. And that actually means making different choices than a race against China, where apparently everything is on the table and everything becomes fair game. It's a race to the bottom, effectively. So you're exactly right that what we need to look at is a much more, I think, grounded, research-led view of how AI can function. And in many ways, looking at the bigger picture, who is this designed to serve?

20:39And who is bearing the costs right now? Because we're seeing in many places, not just in the United States, across Europe, India, you name it, China's the same. It's actually marginalized communities who are bearing the greatest downsides of these new AI infrastructures. So it's going to affect individuals, it will affect communities, It'll affect politics. It'll affect the environment. So that's why I think this really is a societal set of questions, which means it has to be pursued democratically. And I'm not understanding even the lack of foresight there because it's harming the community and eventually will come back to haunt the plants there too.

21:17Right. So, I mean, I think in many ways this is the problem of a race dynamic. People will think short term and they're not thinking big picture. And in some ways, I think that's what communities are really asking for now. They're saying, we care about where we live. We care about our community. We need our water supplies. We don't want to be paying 10x what we currently pay for electricity. All of these questions, economic, environmental, individual, are coming to a head. And I think that is pushing a different conversation about AI that is well overdue. And I think it's really important that you did mention there are choices that can be made.

21:56So again, there are different futures that can be pursued. It is absolutely not zero sum from where these data centers are located, but start with maybe not in deserts. The architecture of chips, the energy choices. I mean, had the U.S. perhaps not shut down so many nuclear sites and had renewables been a bigger investment and had data centers been placed in non-water scarce areas, it is possible we wouldn't even need to be having this conversation or it would be very different. We would have a different focus. But how do you look at the race? I think anytime that there is a race, some people end up getting harmed.

22:33But the reality is the geopolitics are real in this moment. There are allies becoming adversaries. It is a stunning moment to be bringing on one of the most consequential technologies of all time in this era of distrust. So how do you even share the narrative that the race is hurting us when many people will view the outcomes of the race as existential depending on who wins? And I think that is already part of the problem of the framing. Who wins? Who wins what? I think that really has to be asked in a much more honest way. And I think this geopolitical framing is being abused in many ways. Again, as part of my research, you know, I've traveled to a lot of labs where people are making models.

23:27I've traveled to the data centers and talked to people who are sort of installing these systems around the world. People feel as though they have to be in this highly competitive framework. But we're going to need to find pathways where people can actually share information, collaborate. We're seeing a whole new conversation spin up around open sourcing and models. There is a desire for people, particularly in the scientific community, to think about how do we build these for everybody to be better, to be more sustainable, to be less socially harmful. Those conversations are getting shut down by the emphasis on national security.

24:03So I think you're starting to see, particularly right now, researchers try to have a different conversation because they're seeing the real downside of this all or nothing thinking. And what's fascinating is that in many ways, AI systems that don't work properly for people or a planet that continues to burn isn't of itself a national security crisis. Exactly. Most militaries have listed climate change as a massive threat or a threat multiplier from their drones don't take off or they crash or they can't control them. They have to move bases. So all of these things still, even if you can only look through this unilateral lens of national security, this will come back to haunt you.

24:46And if you have models that don't work well for certain communities, I mean, whichever country builds the models that do and are less harmful, that is going to be, they'll have a bigger market share. So all of the, it's this kind of roundabout cycle and it's really hard to find the logic in this moment. Well, I think it's because it's coming at a time where the sort of political thinking is really moving towards technological sovereignty and almost autarky, the idea that you can be a completely independent nation. Now, that's in complete contradiction with the world that we live in, which is extremely interdependent, where supply chains trace around the world and invoke so many parts of every continent in order to produce something as complex as a semiconductor chip.

25:35We are living in a world that is highly interdependent, and that means this idea that you can make everything yourself, that you can be absolutely number one, and you can be the tech-dominant power, which is, of course, the language we're hearing, I think is really missing the deeper underlying realities. So that's the question that we need to be posing, is what is the political vision that in many cases is driving this era in AI? And how does that change? I mean, for somebody that sits in industry and academia, and you also advise world leaders and politicians, is there overlap in the visions or how do you see the different visions?