DeepDive: AI and the Environment

I'm 60% of the world's energy still comes from fossil fuels. And whilst we do have an uptick and an increase in renewables, what we also have is an uptick and an increase in demand. An AI is often shown as one of those things forcing that demand. So if we've got renewables coming online at a certain pace, an AI beyond that pace, which I believe it is, we've got a problem. We're not going to be able to get off fossil fuels because we're not able to replace fossil fuels with renewables. So it's definitely an issue. That was Hannah Smith, Director of Operations for the Green Web Foundation. And welcome to the Machine Ethics podcast. This is episode 100, a deep dive on AI and the environment. This is an episode I've been hoping to produce for a long time. It was recorded across February in 2025, as we gathered four fantastic speakers to help us demystify AI, his impacts and its opportunities for the environment. AI fundamentally is using compute to produce something. And at the core, it's not different from traditional software. I'd say the reason I even have my position is because of the vastness of the latest models that are powering modern AI, so-called generative AI models, large language models, that just use orders of magnitude more compute than any other software in the past has ever done. That was Boris Gamazacicov, Head of AI Sustainability at Salesforce. To produce this amount of processing, AI requires resources. Here's Hannah to explain. The key difference with AI versus another digital technology is it's like 10 times more, 10 times more resources, 10 times more impact. And there are three really kind of crucial areas where there are big environmental impacts. And of course there's energy, which is the one that most people are very aware of. We need electricity to manufacture and power these things. And that manufacture part is at least a significant as the usage part, which again a lot of people don't realise they sort of think if I'm using the thing, that's all they're seeing, but somewhere along the line, the servers, the cables, the devices, they all had to get made as well. And that takes a lot of energy, electricity. And then kind of moving on from that part around the manufacture of things, all our digital stuff, the stuff that you and I use, but also the stuff that powers AI in the background, the chips, the servers, all this stuff, that all has to be built from something as well. So this is where we sort of think about this other environmental impact of the rare raw materials that need to be mined and processed in order to manufacture this stuff. And there's a lot of really sort of dodgy stuff when you start to look at the amount of rare raw materials that are available in the world and when we're predicted to kind of run out of them. And we are predicted to run out of quite a few key ones, like we already know, researchers are forecasting that there aren't enough resources to give everyone an equitable access to technology. That's a really big problem. And then the third thing is around water, which is a really, really big factor in the environmental impact of digital and AI. Data center is needed to manufacture stuff. So if you look at the processes of manufacturing, semiconductor chips and things like that, it needs vast amounts of incredibly pure water, which is not that easy to always find. And then if we also look at the running of data centers as well, data centers typically use quite a lot of water and fresh water to cool the servers. So we're also seeing a lot of pushback from data centers being built in drought-prone areas. And they're kind of taking the fresh water away from local populations. And as one of the things we know with climate change is water stress is more of a factor. If somewhere has already dried, chances are it's going to be getting drier. And data centers tend to be built near population centers as well. So also a bit of a problem. Google actually just released a lifecycle assessment just a few weeks ago on their TPU, which is essentially their version of the GPU, so they're AI hardware. And they found that the embodied carbon, so this is just from kind of the carbon perspective, they found that the embodied carbon, so the manufacturing of the hardware and of the data center. It's around 10 to 20% I believe. So roughly in that order of magnitude, and then the rest is the energy, the actual ongoing operational usage, so that the actual kind of compute provided rather than embodied. So yeah, it really is the ongoing computation for training and inference of these AI models that seems to be dominating. Again, this is just the perspective from carbon emissions and I do think more research needs to be done on other impacts, but I think it's it is useful to kind of put things in perspective. Will Alpine, co-founder of Enabled Emissions Campaign, explains there's more than just the direct carbon impact of the technology. You have two different types of impact. You have the direct impacts and the indirect impacts and it's a really important distinction to draw. So you can think of direct impacts as what goes into the tool, for example, how that tools made or what that tool might run on. Then the indirect impacts would be how a tool is being used. It's important to note that these types of impacts are far less understood and they have much higher impact. On the indirect side, you really need to think about how the tool is used and so AI can either accelerate or mitigate climate harm. It's especially relevant in high emission industries such as the oil and gas sector. What I have seen is that one of the biggest use cases of AI today is actually to accelerate fossil fuel expansion. So that would be exploration and production. AI helps process staggering amounts of seismic data and operational data and make it really easy for oil and gas companies to stay effective with their work, which really widens the gap between low and high carbon energy sources and it really undermines all of the good sustainability work that's happening globally. So concerned with these direct impacts, Mel Hogan, fellow podcaster and associate professor of film and media at Queen's University Canada talks about AI in our current political climate. The way we think of AI right now, and I'm going to put it in quotes, in air quotes, I guess for audio purposes, but I think it's really important to put AI sort of in those quotes because it's a marketing term. So if we take AI in quotes, like the big AI, the generative AI that everyone is talking about since 2022, it's very hard to just talk about the environmental impacts without talking about the ownership of those means of computational production. So already we see in the US in particular with the recent elections, like a quick and efficient turn to fascism and through these AI technology, which have a particular politics related to the environment. So we will see more extractivism, we will see more sort of marketing towards really linking energy companies to AI data centers. We will see more like explicit advancements that are to make AI as big as powerful as possible. But what about the potential environmental benefits of AI? Here's Boris again. I think the promise of AI is vast for a number of environmental and other challenges around us, you know, everything from simplifying complex sets of data, figuring out, you know, best actions to take from a complex set of starting points, speeding up science, speeding up potentially material discovery. So new materials that could substitute for some of these critical materials that are out there, you know, general optimization, automation, it's, you know, AI is such a general purpose technology. I think there's almost an endless array of opportunities. Both Will and Boris point out that not all AI is quite in their usefulness for preserving the environment. There are two different types of AI and I think they're actually being conflated in this discussion. On one hand, you have the analytical AI and then on the second, you have a generative or a gentick AI. And so the type of AI that's really helping the climate is actually quite different from the type of AI that's consuming a lot of energy. So analytical AI is really high accuracy and low energy requirements and that's what's really helping the climate crisis. You can think stabilizing the grid or integrating renewable energy into the grid. And then you have generative or a gentick AI, which is not accurate, but it has really high energy consumption requirements and it's still at a speculative stage and you can think of chat, GPT or any of the associate applications of that. And very often what we're seeing is the AI applications that are most well suited for solving environmental challenges are not the same ones that are the most energy and intense general purpose models. At the AI Action Summit at an event, a researcher noted that for one application, I don't remember exactly what it was, but they found that a thousand parameter model they developed was more effective than the leading kind of general trillion parameter model that is out there. So that is a really great example of not necessarily needing to throw the largest hammer at every problem, but instead creating and applying scalples, so to speak, to certain issues. Hannah is much less optimistic that AI can help. - We talk a lot, and we hear a lot from people say, AI is a climate crisis. AI is good for sustainability. There are no studies. There is no evidence to prove that. What's so ever that is wishful thinking. That is hope that AI will help us with climate change. And as I mentioned in some research situations or perhaps in some situations where you're dealing with large data models, yeah, in the hands of total experts, I think maybe there is a five or 10% promise there that might come through. I don't think AI can ever be a climate change solution if it's used to speed up the efficiency with which we get fossil fuels out of the ground. But just that alone doesn't make any sense. We've got to think about technology and we've got to think about what it's used for. And when you look at the promise of AI in terms of helping to solve climate change, and you compare that against extraction of fossil fuels and making that faster and more efficient, it's not going to just don't stack up. They really, I really struggled to kind of understand how somebody could honestly be thinking that, as a genuine thing. And of course, the other thing I would say on this point is we already have the solutions to climate change. We know what we need to do. We don't need AI, as I say, in the vast majority of cases. I think five or 10% yes, maybe there is some genuine help there. But the vast majority of the problem around climate change is people and politics. And I don't see how AI has a place in moving that along. So I think that's another facet to think about. And a question when we're talking about this promise of AI as a sustainability solution. - Mel seems to agree with Hannah and demonstrates the politics inherent in generative AI. - When generative AI became really popular, I was kind of impressed at how quickly the critiques in and around the environment came out. I'm thinking of shallow-ren who talked about the water usage of AI data centers in particular, did really amazing impressive calculations about all that or Sasha Luchoni and others measured the computational power and energy required to train large language models. So those things were assessed really quickly. I think where we're at now, in part because of the US elections, but also just like a global turn to fascism, we need to be more interested in bigger questions about how the labor of AI gets outsourced to countries that are being recolonized or expansion of colonialism and imperialism through the human labor behind AI. And then with that, all these kinds of sacrifice zones that are at the service of AI expansion. So I think that's like the bigger frame, rather than measuring sort of, which I think are really important, but rather than just measuring sort of the immediate impacts in terms of energy, water, land, et cetera. But I think there's like a kind of political economy also argument forming around this. - Can AI be used sustainably? And how can we get there? Boris and Will explain. - I think everyone really has a role to play in making AI sustainable. I do believe that AI can be sustainable. Just wanna underline that. Right now, there is a race to create the biggest, most powerful model and there hasn't really been in a lot of emphasis on thinking about the sustainability of these solutions. And so I hope that folks can realize that they do have a role to play to whether they are a customer of these companies as an enterprise or an individual user. I think now is a really important time to use your voice and demand the type of transparency, demand more choice in terms of which models I'm using to power whatever solution I have. And yeah, I think it's really important to come together. And then really demand action. I think there is a really important role for smart regulation in all of this. And I do think that that can be put in place without slowing innovation down. - And I think one of the first things that any practitioner can do would be to choose the right tool for the job. And realize that we don't need a massive data center build-up to address climate change. We actually already have most of the technology we need. What we need to do is say, not use AI as a hammer in search of a nail and really, for example, don't use check GPT as a calculator. To choose the right tool for the job. And oftentimes analytical AI and machine learning could be one of the best tools that you're disposal. - Mel echo is the sentiment that we probably don't need huge models for all tasks. - The kind of computational power that you are told you require. And I think deep seek has challenged this. And I think there have been other models that have run even more sort of efficiently cheaper less energy than deep seek. Like I think, you know, people are kind of challenging that notion that those big tech companies were the only ones that were able to do AI at the scale. Sam Altman was asking for like $7 trillion of investment in nuclear energy and so on to power this thing that required all of that. So that's been debunked a little bit. It's incredible that it didn't pop the AI bubble. But I think it has in some sense subverted the idea that it requires that much power. And in both senses of the meaning of power. So I think that that's interesting and what it does is maybe can return us to other uses of large language models, machine learning, even neural networks, probably to a certain extent, and how those things can work for various specific tasks. Like maybe more tedious tasks. Like maybe the truly unpleasant things that humans, you know, that human labor could be sort of like a little bit wasted doing. - So what exactly can we do today? - I think if you're a developer and you're building AI or you are implementing AI in some ways, there are things you can think about with regards to where the energy is coming from for your tools and how efficient you're making those tools as well. Now I work a lot with emissions reporting and trying to get numbers estimates from people who are as to, okay, what does this AI tool actually need in terms of energy and what are the environmental impacts? I would ask any AI developer to be thinking about that and trying to measure and publicly show what they think those energy estimates are. And there's some really cool work coming out from hugging face at the moment around sort of trying to have energy ratings on AI. And I do think we're going to be seeing more of that in the future. So it's about making things efficient but there are also choices you can make with regards to what hosting companies you use. You can pick greener hosting companies. And the best ideal solution is to pick hosting companies that live in or that are cited in regions that already are 100% renewable. And the other things that you can do is think about when you're running AI for training purposes. Looking for times in the grid when the energy demand is lower. There is a correlation between the energy demand being lower and the carbon intensity, which is a measure of how dirty the electricity is. There's a relationship between energy demand being lower and the energy being cleaner because obviously you're using your renewables to their maximum capacity. So there is also a role that developers can play in being mindful and thinking about those things. But the biggest thing is please publish your data. That is the most useful thing collectively that we could shift and change within this industry at the moment. Businesses really need to be transparent and held accountable such as disclosing the AI-related emissions as well as any of the risks of this AI boom on their business or operations or to their shareholders. And then companies are also really-- they need to be held accountable for the climate positioning and the integrity of any corporate sustainability claims. As many of us know, a lot of tech companies have been backsliding on their corporate commitments because of the AI boom, but that's really unfortunate. And I think fundamentally what we need is policy in place. So what governance does a company have in place? What landscape is holding them accountable? One thing-- that employees can do is start to push their company to include environmental considerations into their responsible AI practices. And you could actually, you could start really tactically and say file AI safety violations when you see harmful uses that don't align with climate science. And you can continue to push your employer and just ask why are you not including this because in responsible AI, it must account for harms to people and planet, right? If I was a designer or a company, I would probably say like, I'm an AI free company, and I would like certify that in some way as the selling point rather than say, here's how you can use AI well or even responsibly or ethically, I think we're not going in that direction and be very hard to make a case for that in its current like use and the way it's owned in the way that it's like framed for us in the ways that were asked to invest like emotionally, psychologically, but also like financially, if you think of all those sort of tax dollars that are going to go to fund infrastructure for these things, I think I'm still sort of on the side of opting out, resisting and pushing back. I am just not convinced that this is truly genuinely adding anything to anything. Boris and Will, above working on projects, to help improve the situation. We're seeing more and more regulation where companies are having to disclose their carbon footprint and start to disclose plans to reduce that footprint. If they're procuring AI either for internal uses or for external facing products, that would be purchasing a fleet vehicle and not knowing how much gas it would consume. There's a big liability on these companies, scope three emissions right now. It all really starts with transparency. That's why we were excited two weeks ago to launch this initiative called AI Energy Score. This was created in partnership with Hugging Face, Cohe here and Carnegie Mellon, where we put forth a framework for measuring the inference energy of AI models to be able to start having this informed conversation on what model for what task makes the most sense. We have a lot of data out there on model performance, but we don't have the data on what that trade off is in terms of the energy and all those other impacts that I started mentioning. Currently I'm a climate co-founder of an advocacy and accountability work. We're quite new. We're pushing to align technology use with climate science. This new organization that we're creating is the result of many years of work. Four of those were inside Microsoft, really pushing to align its business activities with its stated support for climate science. The realization came from the fact that one of the biggest use cases of AI today is the by the oil and gas industry to find an extract more oil. I take this really personally because I was involved in building the platforms that are being used by these oil and gas companies to expand production. What we realized is that the additional emissions from just a few of these deals were washing away all of the good we were doing with our green software engineering initiatives or even many times over to Microsoft's own operational carbon footprint. That would include all scopes, including data centers. First and foremost, we're a policy advocacy organization. We're trying to push for policy that aligns use of technology with climate science. We're also educating the public about the full systemic impacts of AI. We're building a coalition and really mobilizing partners across tech and climate and investors. You can learn more at www.enabledemissions.com. To finish, here are some final thoughts from our contributors. We have to critically engage with AI and ask what it's for and what are the long term implications of us replacing our creativity and intelligence with a machine that's not controlled by us. Because ultimately, I think that's the big issue here. AI right now might feel like a bit of a holiday. The tools are free. Isn't this a wonderful boost for humanity? But we've got to critically ask ourselves, who controls this technology? What are their goals? Their goals aren't to serve humanity. We can see that the big tech companies, their goals are serving their shareholders. They might have in the past pretend they'd or really thought that they were trying to make humanity better. But I think we can very, very clearly see now that that is just absolutely not the case. It's about them making money. So what are we prepared to give away for this perceived efficiency? Now, those are very philosophical questions. But I do think people need to stop and think about what they're doing and what problems we're storing up by the future by investing too much of our livelihoods or our processes into these things. And I think as most people who probably will be listening to this know, but AI is drawn from data sets that are extractive. So it's just like pulling, lifting things without consent, usually from the internet or from large scale digitization projects. So it's done without the consent of writers and artists and so on. And that goes for a text, but also for images and video. And I think it's become even more obvious in other formats like video and images where you can really see that it's just like, you know, very derivative from other people's work with no compensation, no consent, right? So it feels to recognize copyright, but also feels to recognize that like AI does not label itself. AI does not understand meaning that there's a lot of human labor and usually outsourced, you know, to kind of like these call centers, whether, you know, Asia, Africa, like it's outsourced from the US, the folks working in these centers are often misled. Also don't really have a choice but to work for these kind of companies and often suffer like trauma from the stuff that they have to sort through. So I think that's pretty well documented as well. And then there's the environmental aspect where I'm like, if you know how much water energy it uses and like you don't care. And you have these three levels of not caring already, then like go for it. But like I think teaching people to care and teaching people like about the impacts and the like complete unsustainability on three or more levels, all of it is interconnected. And that's that's the problem with AI. I would start everything with the need for more transparency is where we are right now, the leading AI model providers, they're most popular and you know, most widely used and also, you know, most frontier models are essentially black boxes, you know, outside of other issues with that from an environmental perspective. That means when we're using these systems, we don't really understand our environmental impact from a individual that's, you know, problematic for sure from an enterprise that's using AI that can be potentially legally risky. Let's get ourselves out of this false dichotomy discussion between the direct impacts of AI and its potential benefits because that's an overly simplistic view. Focusing on just one of these while understandable is really inadequate for the task at hand. So I need you to choose your issue regarding AI and then zoom out to think about the full system that it's part of. And my conclusion having spent four years trying to make positive change from within is that the only thing that will save us now is policy because technology is just a tool. So if you're in a company, the single biggest thing you can do is to push your employer to use their power to advocate for policy that truly aligns tech with climate science and hold them accountable for promises regarding accountability and transparency. Because traditional sustainability practices such as incremental measurement and carbon reduction are insufficient given the challenges that we face. Welcome to the end of our deep dive episode on AI and the environment. Thank you so much to our speakers Hannah Smith, Mel Hogan, Will Alpine and Boris Gamazacikov for feeding my questions their time and their amazing knowledge. This is our 100th episode of the Machine Ethics podcast. Thank you so much if you've been listening and thank you for joining us if this is the first episode. You can find many more machine-ethics.net from our other deep dive episode on AI and Games. To conversations about science fiction, transhumanism, parenting, AGI, automated cars, building and using AI in your companies, as well as various looks at ethical frameworks, responsible AI and AI ethics in general. If you'd like to find the full interviews from our contributors for this episode, you can find them by joining our Patreon Patreon.com for a slash machine ethics. You could also find more thoughts from me on the Patreon as well as other exclusive content. If you would like to continue and come with me on this journey of AI and its impact on society, then you can follow us on Blue Sky, Machine-ethics.net. Instagram, M/M, machine ethics podcast, YouTube, @machine-ethics, and wherever you find your podcasts. Thank you and I hope to see you at Episode 200. See you then. [Music]