On the environmental costs of generative AI
The environmental costs of generative AI are often discussed either in absolute terms (what are the direct ecological costs?) or relative to Web searches (how many Google searches does one ChatGPT query cost?). I think this is a mistake, even as an approximation, because generative AI affects our other behavior so much.
An example: a little pump on my furnace1 stopped working recently. I discovered this when I heard a strange, continuous noise. With ChatGPT's help, I diagnosed the problem, drove to Lowe's, and fixed it that night. This is very different from how the situation would have played out a couple years ago, when I would have called in a professional.
Generative AI's effects were large along many dimensions here, including environmentally. Service calls and pumps are orders of magnitude more expensive than generative AI conversations.2 There are times to ignore second- and higher-order effects of a decision, but this is not one of them.
I'm not sure which direction the broader, fuzzier considerations weigh in: here I'm pretty sure ChatGPT gave an environmental win, but in other cases it surely leads me to do things that are eco-costlier than I otherwise would have. As we look beyond me and beyond now to the whole future of the economy, I certainly don't know what the net environmental effect will be. But I'm pretty sure that it won't be calculated by looking narrowly at data centers or CPU cycles.
Note: I don't particularly like to write about politics, but I find myself repeating this, and blogs are great for recording thoughts you find yourself repeating. Also, I'm proud of the furnace repair and am now able to brag publicly about it.
It's a condensate pump, for getting rid of water emitted by a high-efficiency furnace.↩
Depending on the details of my guesswork, things other than computer use account for between 99% and 99.99% of the total environmental effects of this event. (And if I had to bet on either "over 99.99%" or "under 99%," I'd take the former immediately.)↩