Climate and AI

I Just Don't See AI Having Much Impact On Emissions, In Either Direction

You may have seen talk about the impact of AI on climate change. There are hopeful articles listing ways that AI might help reduce emissions. At the same time, people fret about the electricity used to train and operate AI models.

People ask me about this. I tell them that both ideas strike me as overblown. For the foreseeable future, I don’t see AI significantly changing the trajectory of climate change, in either direction. On the one hand, the amount of energy used for AI is just not that much on a global scale. On the other, the pace of emissions reduction is primarily limited by our ability to execute on things we already know how to do, rather than our ability to invent clever new things.

Using AI To Reduce Emissions

Don’t worry, I do realize that if we use AI to optimize the grid, it won’t actually involve a humanoid robot pushing buttons on a control panel.

Speculations as to how AI might help reduce greenhouse gas emissions mostly boil down to “AI could make [thing] more efficient, thus reducing that thing’s emissions”. The thing in question might be supply chains, the electrical grid, forest management, heating and cooling, etc.

There is also talk of AI helping make communities more resilient to climate change, e.g. by using weather models to better plan for shifts in rainfall patterns and other impacts.

These proposals tend to have two things in common. First, they’re vague. When someone is writing about a new trend, vagueness is a bad sign. It’s a tell that there’s nothing really happening here – at least, not yet. If AI was having an actual impact on emissions reduction, or there were specific startups or projects promising an impact, all the articles talking about AI and climate change would cite examples¹. The lack of such examples tells us that this is a made-up trend: people aren’t writing about it because they’ve seen it happening, they’re writing about it because they assume it will happen. I mean, obviously AI will have an impact on climate change, because – well, gosh, it’s going to impact everything, right? Spoiler alert: no, in the near term, AI will impact some things but not others and it won’t often be revolutionary, especially with regard to the sorts of physical-world activities that are relevant to greenhouse emissions.

What’s the second common element in proposals for AI-based projects to reduce emissions? To the extent that there’s actual potential in the proposal, that potential generally doesn’t have anything to do with AI. For instance, consider the idea of improving management of the electrical grid. There is absolutely room to better manage the grid, as I touched on in The Electricity Transmission Challenge. But the barriers to improved management are mostly things like “utilities are not incentivized to do it”, “management of the nation’s grid is fragmented across too many separate agencies and operators”, and “it’s a conservative industry where it’s hard to make changes”. Those aren’t problems that you solve with AI.

To put it another way, I strongly suspect that any proposal to improve grid management using AI will look like this:

1. Come up with an efficient plan for using grid resources.

2. Carry out that plan.

AI helps with step 1, but we don’t really need help there; it’s step 2 that has been holding us back.

Of course, AI is hot right now, and so everyone will be looking for excuses to slap an “AI” label on whatever they were doing anyway. Who knows; sometimes this might even succeed, as a matter of salesmanship, in attracting buy-in to a project. As a result, we may see some successful emissions reduction projects that are hyped as incorporating AI; that won’t necessarily mean that AI really was important to the outcome. Conversely, we might see projects which fail because everyone focuses on the AI part and then the project founders on the barriers to action – siting, regulatory approval, etc. – that make so many mitigation projects difficult.

Over time, there will be instances where newer AI techniques allow us to shave off a few percentage points here and there. I just don’t expect it to be a game changer anytime soon.

AI’s Use Of Electricity

From The New York Times, in October:

OpenAI’s ChatGPT exploded onto the scene nearly a year ago, reaching an estimated 100 million users in two months and setting off an A.I. boom. Behind the scenes, the technology relies on thousands of specialized computer chips. And in the coming years, they could consume immense amounts of electricity.

A peer-reviewed analysis published Tuesday lays out some early estimates. In a middle-ground scenario, by 2027 A.I. servers could use between 85 to 134 terawatt hours (Twh) annually. That’s similar to what Argentina, the Netherlands and Sweden each use in a year, and is about 0.5 percent of the world's current electricity use.

It’s worth noting that there is a lot of spin going on here. The first paragraph primes us to think big: ChatGPT “exploded”, there’s an A.I. “boom”, which could consume “immense amounts” of electricity. There’s also an implication that AI uses enormous numbers of chips (“behind the scenes”, no less – how suspicious!), but in fact “thousands of specialized computer chips” is quite a small number compared to the scale of modern data centers.

The second paragraph may leave you with the impression that AI is projected to use as much electricity as three entire countries – Argentina, the Netherlands, and Sweden. But the word “each” means that actually, the projection is only for AI to match one of those countries, not all three. So, why not just say “similar to what Sweden uses in a year”?

Spin aside, let’s dig into the numbers. Weirdly, I couldn’t easily find a figure for global emissions from electricity usage, but the “electricity and heat sector” generated 14.6 Gt of emissions in 2022. Let’s conservatively assume that electricity accounts for most of that, call it 12 Gt. If A.I. in 2027 is using “about 0.5 percent of the world's current electricity use”, that would be 0.06 Gt / year in CO₂ emissions².

By comparison, livestock production generates around 6 Gt of emissions per year. So, if you’re worried about emissions from AI, here are some things you could do, each of which would have the same impact:

• Shut down AI entirely by 2027

• Reduce livestock consumption by 1%

• Make livestock management 1% more efficient

• Reduce methane emissions (e.g. leaks) from the energy sector by 0.5%³

• Use AI to achieve 0.1% in efficiency improvements across the global economy⁴

Even this paints AI in an unfairly negative light. A lot of AI chips go into data centers run by tech giants like Microsoft and Google, which are aggressively pursuing clean sources of power. Not only are these companies signing contracts to purchase electricity from solar, wind, and other clean sources, they’re often paying a premium to be early adopters of technologies such as advanced geothermal power or small modular nuclear reactors, thus helping those technologies achieve commercial viability. So data center electricity is probably cleaner than the overall electricity sector, and the tech giants are actively helping it to become cleaner still. (I’m not sure whether Amazon is contributing to the clean power transition as aggressively as Microsoft and Google.)

Also, to a substantial degree, electricity usage for AI is displacing electricity usage for other purposes. The world has only so much capacity for producing the kinds of advanced chips used for AI training and inference, and it’s running flat out. Every chip used for AI is a chip not used for, say, crypto.

In The Long Run, All Bets Are Off

I started this article by saying:

For the foreseeable future, I don’t see AI significantly changing the trajectory of climate change, in either direction. [emphasis added]

When it comes to AI, it’s not clear how far the “foreseeable future” extends. In the next five years, for sure, there won’t be much impact. We won’t be able to build chips quickly enough to amount to much net additional electricity usage, and neither will we have time to deploy any AI-assisted efficiency improvements at scale.

After 5 years, I still don’t see electricity usage from AI becoming a major concern. If we’re using AI so heavily that it starts to become a significant portion of global energy demand, then we’ve entered the “grander visions of AI” phase (see next paragraph). However, it does seem possible that we will eventually start to see substantial efficiency improvements due to AI. For instance, AIs might help develop new catalysts, enzymes, or genetic enhancements to improve efficiency in food production, chemical processing, and industrial processes. They might reduce soft costs for rooftop solar and other renewable energy projects, by helping developers fill out paperwork and even helping government agencies process it. This could sometimes be game-changing within a specific sector. But it will continue to be less important than finding ways to cut through red tape, get utilities and other incumbents to embrace change, and generally plan and execute better as a society.

If and when the grander visions of AI start to become reality, all bets are off. Maybe we’ll develop hyper-efficient chemical engines that use solar power to 3d-print anything you want using CO₂ captured from the air. Maybe we’ll spend our lives in virtual environments that don’t require us to ever get in a car, let alone an airplane, nor to use more living space than the size of your VR pod. Maybe robots will make mass-produced junk so cheap that we all buy a lot more of it, driving emissions ever higher. Maybe an AI-driven economic boom will leave everyone feeling prosperous enough to be comfortable paying a green premium to drive emissions to zero; or perhaps AI will exacerbate income inequality and political tensions to the point where advancing climate-friendly policies becomes impossible. In my AI blog, I’ve written about how a future with powerful AI will be utterly different from the present day. I don’t think we can predict how that will unfold; at best, we can focus on keeping up with developments, and being prepared to act as needed. “Chance favors the prepared mind”, as the saying goes.

Thanks to Michelle Deatrick for inspiring this post.

1. I’m being a bit glib here. There are instances where some form of AI is being used to in some fashion help with emissions reduction. However, to my knowledge, the contribution of AI is usually not dramatic, and – this is the key point – I’m not aware of examples that relate to the new wave of generative AIs (ChatGPT et al). In other words, I’ve seen nothing that points to the likelihood that AI’s impact on climate change mitigation is about to change dramatically as a result of recent progress in AI.

2. I see some issues with the math used in the source cited by the Times, but I don’t have a better estimate to replace it with. Even if electricity usage for AI turns out to be 5x higher or lower than the estimate, it wouldn’t drastically affect the analysis I’m performing here.

3. The IEA’s Global Methane Tracker 2023 puts methane emissions from coal, oil, and natural gas production at around 125 Mt/year. Over a 20-year horizon, methane has 80 times the warming impact of CO₂, so this is the equivalent of about 10 Gt/year of CO₂. The report goes on to note:

In the oil and gas sector, emissions can be reduced by over 75% by implementing well-known measures such as leak detection and repair programmes and upgrading leaky equipment. In the coal sector, more than half of methane emissions could be cut by making the most of coal mine methane utilisation, or by flaring or oxidation technologies when energy recovery is not viable.

4. Global emissions are around 50 Gt/year. 0.1% of that would be 0.05 Gt/year, roughly equal to the projected emissions from electricity for AI.