Here come the rationalisers

About two years ago, I wrote something about the carbon footprint of generative AI, questioning whether it was a good idea to argue against AI based on its environmental impact. My argument then was basically that, if you say using AI for creative work has an unacceptable carbon footprint, then you're also implying that every piece of creative work has an "acceptable" carbon footprint, however it's produced — and once you do that, you're opening the door to quashing human-produced creative work, whether AI is in the picture or not.

I stand by that argument, but I think it needs a follow-up. Because more and more, lately, I've noticed the pendulum in AI debates swinging too far the other way: lots of people are now saying the environmental impact of generative AI is trivial, particularly compared to the other stuff we all do every day. For example this fellow called Andy Masley has built a little calculator, where you can put in how much you use AI, and it will tell you how much AI contributes to your daily carbon footprint. Spoiler alert: no matter how crazy the usage figures I put in, I really struggled to get it to tell me that AI contributed more than 10% of my carbon footprint. Turns out your car and your steak dinner are much worse for the planet. Have a little Google for "AI carbon footprint calculators" and you'll find dozens more telling you roughly the same thing.

Even among people who tell you candidly that a session with an AI chatbot uses significant energy, some now hedge their advice. Yes, asking ChatGPT "what's the carbon footprint of AI" will use a lot more energy than doing a simple, AI-less Google search for the same thing — because Google only has to go retrieve results from its search index, whereas ChatGPT has to both find the results and write you a response with the answer.

On the other hand, Google's search results are only the first step. You then need to go click on a lot of those different results, which in turn requires retrieving a lot of different websites to load in your browser, and if one of those results includes a 10MB YouTube video, then suddenly the carbon footprint of your Google search is beside the point — because the carbon footprint of answering your question is now much higher. (And if you don't find quite the thing you're looking for with your first Google search, you'll then do another more targeted one, and click some more results, and maybe then another query, and...)

As a result, some people will now tell you that the carbon footprint of answering your question with AI might actually be lower than the traditional noodling around on Google for ten minutes — particularly if you can give your AI a good, detailed, specific prompt upfront. At the very least, once you get rid of all the clicking around and follow-up Google searches, the carbon footprint of answering your question with AI will be broadly similar to having done it the older way.

Likewise, asking Claude Code to build you a website might use several hundred thousand tokens and take an hour or two. Nobody really knows how to quantify the exact carbon cost of that, but even if it's a couple of litres of water and a kilo or two of carbon... Well, a human developer working on that same website over a week would surely drink more than a few litres of water in that time, and use far more than a kilo or two of carbon. This is exactly what I was worried about two years ago: if we're going to start making this kind of comparison, there's no guarantee that human creators will be the most carbon-efficient solution.

There's a bigger problem with thinking about the issue this way, though. Yes, for some complex tasks, a well-trained person using AI in a targeted, thoughtful way might have a lower carbon footprint than someone else doing the same thing manually. But, first of all, that puts the onus on the user rather than the provider: if AI is killing the planet it's because you're using it wrong, not because there's anything inherently bad about the technology.

And second of all, that scenario is clearly an ideal case, rather than the actual reality of how AI is being used. According to OpenAI, ChatGPT has 700 million active users every week, entering about 18 billion prompts — and according to the National Bureau of Economic Research, 24% of those prompts are just seeking information (e.g. "Who was president after Lincoln?"). That's a lot of people using a computationally and environmentally expensive tool to find a piece of information they could have found a lot more cheaply using Google. (Or even, dare to dream, a book.)

So, why the shift? Why are people suddenly trying to convince us that AI's carbon footprint isn't so bad? Part of it, no doubt, comes from the AI companies themselves, who have a vested interest in promoting their products' sustainability, particularly in the face of growing backlash against data centres and their effects on the surrounding environment. But just as often recently, I've heard the argument that AI's carbon footprint isn't so bad from exactly the same people who two years ago were ragging on AI's carbon footprint incessantly. So why have they changed their tune?

I can only speculate, of course, but my bet is that they've started using AI, and realised that they like it. I'm actively encouraged to use AI at work now, so I do, sometimes — and listen, it's great at speeding up tedious tasks, or even doing them for you altogether. When you get to the end of the day feeling buried under your to-do list, and Claude can cross off three things in the last half hour, you feel an enormous amount of stress lift. And unlike doing a job manually over several hours or days, you often feel more confident in the result, rightly or wrongly, because AI is always confident in its own answer. Sometimes AI even flatters you by telling you that you asked it a great question, and you get a little rush of silly pride. AI appeals to innate human laziness and venality, doing your work for you and then telling you how great you are for taking the credit. Once you have access to that every day, why would you ever want to give it up?

I'm sure fear comes into it too: because if you're a business, and you think all your competitors are using AI to increase productivity, then you feel like you have to jump on the bandwagon or risk being left behind. And depending on your business, you might also find that AI seems to do better work, on some tasks, than any of the humans you employ — at which point, from a commercial perspective, it feels crazy not to take advantage.

Either way, when you started off saying AI was bad because of its environmental impact, and now realise you're actually quite attached to using it, you have no choice but to rationalise the environmental impact away — hence all the newly minted AI apologists, scolding you for eating steak instead. That might make those individual people feel better day-to-day, but it's a disaster for the planet. Because once the idea is out there that any one person's AI use is trivial, particularly if that idea is coming from people who are otherwise noted for their pro-sustainability views, then more people are going to start using it for more things. In aggregate, all those little trivial queries add up to an enormous drain on resources.

Note that this isn't quite the same argument I worried about two years ago. I'm still not suggesting we should measure the carbon footprint of AI doing a specific type of work, and use that data to decide whether AI should be doing the work to begin with. Like I said two years ago, purely on the merits of carbon footprint, if you compare the cost of coding one website or writing one article to the carbon footprint of a human doing the same thing, you might find that humans are actually the "worse" option. Also, despite what anyone might tell you, it's not really possible to know the exact carbon footprint of writing an article, whether it's an AI doing it or a human, so the comparison is kind of meaningless anyway.

The argument I'm making here is much broader, about the carbon footprint of AI in aggregate. What I'm saying is, it doesn't matter whether AI can do this or that task with a negligible carbon footprint, because overall the carbon footprint of making AI available for that task is untenable. You can't separate out your own, responsible, efficient AI use — optimising prompts and only using AI when absolutely necessary — from the hundreds of millions of other people asking it when Lincoln was born or what they'd look like as a dog. Make ChatGPT and Claude available to anyone with a credit card, and the world will burn. (It already is literally burning.)

I'm not being a luddite here; I'm not saying we should get rid of AI altogether. There are people out there using it for incredible things, designing new drugs and improving cancer screening and making better epidemiological models, and they should absolutely keep doing that. But collectively, as a planet, we need to regulate who is allowed to use AI, and for what — and we need to do so not with glib back-of-the-envelope calculations about whether a particular prompt released more carbon than eating a steak, but with a priori, ethical reasoning about which outcomes are worth the obvious costs of the technology.

No doubt that will be difficult, because it will limit how much money the AI companies can make, and no doubt there's room for significant debate about which outcomes are ethically defensible. But if the cost is annual wildfires engulfing every continent, I think we'll pretty soon come to the agreement that making your email sound friendlier is not worth that cost — no matter the carbon footprint of those individual prompts.