The climate impacts and ethical considerations of AI use

Community
1

It is not possible to use AI and LLM tools without endangering the Julia community. I understand that the question is geared towards the online, programming community (such as this forum). But, let’s be real here for a minute because we all must live in the real world, with real consequences. It is not possible to separate the technical pros and cons from the ethical pros and cons.

I am a member of this community. When developers clear cut and bulldozed ~100 acres (~0.4 km^2) of forest near my house to build a brand new data center, that harms me. When they clear cut another 100 acres across the street to build a new high-voltage power transfer station, that harms me. The unfortunate state of affairs is that users of LLM tools and consumers of AI products and providing the market leading to the data center boom.

The infrastructure that the technology relies on has an immediate and direct negative impact on the communities in which they are built. And the methods with which they are powered and cooled have a delayed, long-lasting, and potentially irreversible effect on the environment for everyone.

In the United States where the most of the “brand name” models are trained (I presume), roughly 60% of the electricity comes directly from fossil fuels. Even though more renewable sources are being built, the total amount of fossil fuels being burned is increasing, almost certainly to account for the explosion in data center power use.

The Southwest United States is has been locked in a mega-drought for more than a decade, and it’s spreading. Europe is seeing a drought and water crisis on a scale that is unprecedented. Africa has had repeated, extended droughts for years. It is likely already too late to hold the world below the 1.5 C target after the world hit 1.4 C above average. Data centers didn’t start the fire, but they are certainly adding more fuel.

And for the big tech companies, that fuel is money. When users buy subscriptions to models, when companies pay AI-advertising companies for their service, when Wall Street investors buy into an IPO, all that money fuels the data center boom. The same data center boom that is carving up landscapes, upending established communities, and dumping ever more CO2 into the atmosphere.

It is impossible to know what the Earth will look like in 2050, but I think we all need to step back from the short-term glimmer of shiny new technology and try to look at the long-term effects of its unbridled growth.

PLEASE NOTE: I am not posting this to call out or demean anyone. Everyone different jobs, roles, pressures, and life circumstances that feed into every decision they make, including myself. We all do the best we can with what we have.

Just try to take the cosmic perspective every once and a while.

2

I’m also absolutely incensed about the climate disaster unfolding before our eyes, the wealth inequality, and the ever-increasing drive towards unsustainable growth. AI datacenters do seem to be the perfect encapsulation of all three.

But at the same time, as I was sitting in the Sustainability in Computational Science and Engineering JuliaCon keynote 7000km from home, it struck me how it’s not terribly unlike air travel. I know multiple people who have personally sworn off air travel — and for good reason. I completely respect and understand their decisions. Personally, I’m limiting my own air travel… but it really is a powerful technology! When I consider a plane trip I do weigh the possible benefits against not just the monetary and time tradeoffs but the environmental externalities as well. I think private jets are unconscionable; if I ran the world I’d ban them entirely.

I’ve also thought about how one of the reasons I got into Julia (and open source in general) was to help others avoid wasting their own time and resources on things I’ve already solved.

And indeed, a well-designed and accessible open source project is a wonderful outlet for AI effort! Folks shouldn’t be individually having AI build a million proprietary one-offs over and over again; that’s like the private jet of the air travel world. But I see huge benefits to having these AI agents build powerful open source systems that can continue rising our collective capabilities — capabilities that I’m also convinced we need to fight the climate catastrophe.

3

I am very happy with your contribution, and I fully agree.

Using AI tools is like eating meat: Bad for the environment, but many of us do it. And this would not be a big problem, if we would do it in a limited way. Like once per week.

I am very much in favor of the doughnut economy: we should develop the economy so that there are no very poor people, while also limiting wealth and resource spending to a level that is sustainable for the earth.

My concrete issue with AI tools: I do not know how much tokens I burn if I use a Claude pro subscription. I do not know how much CO2 emissions this causes. So a first step would be an obligation for the AI industry to provide this information to the consumer.

This will not happen in the US under the current administration, but still worth fighting for it. And AI is not only used in the US.

4

That’s a big problem. No question about it. However, one could say the same about using computers in general. From my point of view, those kinds of questions should be directed toward regulators and energy providers rather than to the end users.

5

You can’t blame individual consumers of AI for trying to be more productive. It’s the same as with cloud technology, we never asked for that either, and yet it was still forced upon us. If you want to stop AI datacentres from consuming too many precious resources, you need to push for more regulation. We can’t stop the progress, but we can reduce the speed at which resources are consumed.

6

The most effective would indeed be government regulation, but you can definitely blame individual users for pouring petrol onto the fire.
Ironically enough what I think will actually happen is that at some point investor money will run out so the big AI companies will have to stop subsidising token costs, which will dramatically lower demand.

7

Individuals make up the collective demand, and that is just as important as the suppliers and the regulators, if not moreso because consumer sentiment heavily affects who end up as the suppliers and regulators. I seriously doubt AI optimists would vote for stronger regulation, or AI skeptics would jump at Anthropic’s IPO. Suppliers spend billions on public relations because their consumers pay it back in spades.

8

This is a very apt analogy. Air travel is (IIRC) very inefficient compared even to driving solo, and yet it does open up the entire world for ordinary folks to discover. So where is that line to draw? (rhetorical question!)

Another good analogy. Especially because I think a lot people consider the ethical arguments around meat eating, and I think using AI deserves the same amount of thought.

I am not blaming anyone individual for doing what they need to make it in this world. But it is indisputable that the collective effect of millions of users is what drives the market. The productivity argument strikes me as identical to farming. I certainly do not blame individual farmers for using petroleum-based fertilizers and pesticides on their crops to increase productivity to make a living and feed people. However, it is also impossible to ignore the environmental effect of the tools/chemicals behind that productivity.

I did indeed vote for my local candidate who opposes new data center construction. But, at least in the United States, those with the biggest wallets cast the most votes. Also, part of pushing for more regulation is trying to get more people on your side and doing things like making forum posts. :nerd_face:

Many of the big tech CEOs paint their view of an AI-dominated future as “inevitable”, but its not. It’s just that the more that something is repeated the more believable it becomes. This is visible in misinformation campaigns that attempt to sway people’s minds. The pendulum of progress always goes both ways.

9

I am a heavy user of agentic AI coding, if I had to pull a number from the increasingly CO2 laden air, I might guess that I’m in the top 5% of usage for subscriptions to Anthropics models, maybe even top 1%.

I’ve done some back of the napkin calculations and I’ve estimated that if you assume the worst case for electricity generation (all dirty coal, no renewables), token usage, caching, batching, etc. that using 100% of a Claude Code Max 20x subscription every week for a year might be equivalent to a single round trip from the Pacific Northwest, USA where I’m located to Mainz, Germany, and back. (I did not attend the conference, but I would have enjoyed doing so, I’m sure)

If you use some more forgiving assumptions, it could take on the order of 10-20 years of AI use at this level to reach the equivalent radiative forcing of one international air flight.

I’m not trying to vilify air travel as I understand that this is a global community and that there are real tangible benefits when people gather together, in full flesh form, in order to meet and exchange ideas. I understand that there are experiences and exchanges which happen in this form that are difficult or maybe impossible to recreate without that physical proximity.

But, it’s also perhaps important to note that the electrical usage from data centers can be made via renewable sources, can be made more environmentally friendly. The same is likely much more difficult in the case of jet fuel, pretty much impossible.

So I think it’s pretty important to keep this kind of comparison in mind, if one is feeling judgmental towards people who use AI tools for coding.

10

Man, chill out a bit, you’re overreacting!

11

OH! Am I? :joy:

I honestly don’t feel like I was, I was rather just trying to add to the conversation. I wasn’t feeling specifically called out or anything, but I was attempting to honestly place myself within the spectrum of AI users and I do understand and consider that use of technology like this is not without cost.

My quoting of @mbauman was to add some numbers to their thoughts, because I thought it was a really great insight, and was actually exactly what I had in mind when I read the topic title.

So, no worries, all chill here, hope you are feeling the same. :slight_smile:

12

If AI truly brings productivity gains there is nothing you can do to stop it. That has nothing to do with what the AI gods or big tech are telling us, it happens because we have capitalism. Those who do the same work with less resources simply push the others out of the market because they can undercut you.

13

At least for me, I truly harbor no judgement for those who use AI to help make their way in the world (except those who use it nefariously, obviously).

Interesting! It also kind of illustrates my point about the collective usage. If your numbers are reasonable, combining that with the estimated 19 million Claude Code users works out to about 1-2 million international flights per year.

That’s a drop in the bucket though since my understanding is that the vast majority of the energy consumption comes from the training steps, not as much the inference steps. Correct me if I’m wrong though!

Sure, it can be more renewable, but currently it is not in the US. As far as environmental effects go, it does not matter what could be in the future as much as what is happening now.

14

Based on 2024 data, there were about 250 million round trip long haul flights of the type I use for comparison, so if all of the ~20 million Claude Code users were at the maximum usage of the highest subscription term (they aren’t), that would equate to in the ballpark of 0.5% - 1% of the aviation impact.

In my calculations I included the amortized cost of training, based on estimates from published figures for comparable models, so the training cost is included along with inference time costs in my estimates. I want to be clear that these are estimates and not intended to be precise, just order of magnitude.

15

This is false. Going back to your idea of regulation, there was a time when the most productive companies forced people to work 12-14 hours a day. Eventually people protested and got fair labor laws passed. There was a time in America when the most productive coal mines also had thousands dying of black lung, then more laws were introduced (sadly, this trend is reversing for miners).

AI is no different than other technologies really. It comes on the scene, changes everything, and then people need to deal with the good and the bad. AI may not go away – coal mines still exist after all – but it may be in a very different context than it currently is.

16

If anyone is concerned about climate change and wants to protest, I’d protest the endless warfare around the world and the funding of massive military programs (my home country being the worst offender here). The CO2 impact of the US military is equivalent to somewhere in the order of several BILLION Claude Code users.

17

AI adoption has scaled to the point this isn’t true anymore.

PEDs are also productive.

The main issue with these arguments is lumping resources into a collective global pool that doesn’t exist; it’s basically like arguing we can solve world hunger overnight simply because we technically have made enough food. The problem is data centers are concentrated in certain populated areas for the associated infrastructure and climate, and that adds up to a significant local burden because the rest of the world can’t just teleport fresh water and electricity there.

18

1% of all flight is pretty large considering that Claude Code users are only one piece of the user puzzle considering there are also folks using models from Google, Meta, OpenAI, xAI, and the like for coding. And my wager is that coding uses are dwarfed by things like AI video generation for social media platforms.

19

This may be true, but is entirely beside the point and not relevant to the discussion at hand.

20

And those international flights are also eclipsed by people who drive a car to commute to work, or people who play a video game on a modern console, or who … the list goes on.

With my estimated numbers, a year of MAX Claude code is the same as about 5% of the impact of the average emissions from driving a vehicle assuming the average efficiency (~25mpg) and average mileage (~13,500 miles) as quoted from the EPA figures.