On this week’s episode of The World Unpacked, Jon Bateman spoke with independent AI researcher Andy Masley about the controversy over data centers in the United States. A portion of their conversation, which has been edited for clarity, is below.

Jon Bateman: What is your bottom-line explanation for the average person who wants to know whether their personal use of an AI chatbot is a significant part of their environmental footprint?

Andy Masley: I’m basically not aware of any major researcher who’s finding that the cost of individual AI chatbots is adding up to something that we should worry about, if we place it in the context of all the other ways that we use energy and water.

I’ll start with water, because I find that that’s something that people worry [most] about. There’s this really popular statistic that one individual chatbot prompt can use an entire bottle of water. The researcher who came up with that statistic—who had reported it to the Washington Post, which popularized it—no longer agrees with the statistic at all. At the time, he thought it was a reasonable estimate. But we now know a lot more about what the models at the time were like and what current models are like. The current models use probably something like thirty times less water per prompt than that. But this specific statistic has stuck. 

This can sound like a lot in a vacuum. I think it’s very shocking for people to hear that their computers use any resources at all. But if instead you ask, “How many AI prompts would it take to raise my personal water footprint by 1 percent?,” the average person’s chatbot prompts probably would require something like 200 to 1,000 prompts to raise that person’s emissions and water use by 1 percent specifically. And this is partly due to the fact that computing is incredibly resource-efficient across most things, so it very rarely adds up to being a big part of our emissions and our use.

Jon Bateman: What’s another thing in the typical American’s life that might be in the same order of magnitude in terms of water or energy use as chatbot use?

Andy Masley: I usually like to compare [AI chatbot use] to using a microwave for a few seconds at a time. If you have a whole conversation, maybe that’s a minute of microwave [use]. That’s definitely not nothing, but I would personally use a microwave for one minute if it could generate this gigantic, useful conversation for me.

Jon Bateman: You mentioned that other forms of computer use also use water and power—sending an email, doing a Google search, storing an image in the cloud, watching Netflix, listening to this podcast. How significant is any of that compared to AI use?

Andy Masley: AI use is moderately resource-intensive by the standards of other things we do online specifically. But the thing is that other things that we do online are just such an incredibly small part of our emissions from the get-go that even a large multiple of those can still be very small.

One example is that an analog watch uses one-millionth the power of a digital clock. If you buy a digital clock, you’re multiplying your time-telling energy use by a factor of a million. Does that tell us anything about how bad it is to buy a digital clock? It doesn’t, because even though the analog watch is so much more efficient, these both round to zero in an absolute sense. It’s a little ambiguous exactly how much, say, Netflix-watching uses in comparison to AI, because Netflix is actually very efficient. Best I can tell right now is it might be something like ten to twenty minutes of Netflix for every AI prompt. If you look at the resources of running your laptop or your computer, then AI actually becomes like running a few minutes of your laptop normally.

So if I met a friend who wanted to reduce emissions as much as possible, and to do that, he ended his laptop use three and a half minutes early every day, I would think, “This is so hopeless as a way of affecting the climate. It’s going get zeroed out by all these other things you do.” There are other changes you can make instead that are much better.

Jon Bateman: Where’s the data coming from?

Andy Masley: There are a bunch of independent experts that I personally defer to, partly because they’re much more critical of AI’s use than me. One organization called EcoLogits estimates its best guesses of AI energy and water use based on what it knows about the software, the hardware, the grid that data centers are built on, and things like that. It’s really not trying to defend AI’s resource use . . . and the numbers it uses mostly converge on the same estimates from a lot of other sources as well. I don’t really try to update too much on any one source.

Watch the full episode on YouTube or subscribe to The World Unpacked.