I'm not so sure about the quality of this answer. It seems to be largely based on "well, I'm going to assert X doesn't use a lot of water, because Y uses even more water". Anyone who's been paying attention, already knows Y's use of water is a huge problem, so the assertion isn't effective.

The paper "Li et al 2024" was posted to arXiv in Oct 2023, so that citation is incorrect. Also it does nothing to support the claim that "a 100-word email [using GPT-4] requires 519 milliliters of water", so where did WaPo get that information from? (See my answer for more information; I'm citing the same sources.)

@EndAntisemiticHate Everything is relative. If my existence causes the consumption of thousands of units of a resource per day, I woulds be confident saying that <1 unit of that resource is not a lot. Additionally, discussion of water consumption is usually in the context of trying to reduce water shortages. If someone wants to argue that 1 liter of water is a lot and therefore we should care about it, I'd argue that we shouldn't care about it since we should care about other things instead, e.g. the meat industry and rampant consumerism in developed countries.

This answer bickers about the vague definition of "lots". Focussing in how much, and then comparing it to beef and rice and letting the reader decide would be more productive. (I expect a meal to take much more resources than an email, so your comparison seems a little empty.)

@EndAntisemiticHate: It seems to be largely based on "well, I'm going to assert X doesn't use a lot of water, because Y uses even more water". I never made any such argument. Please don't put words in my mouth or twist my argument. Try to read what I wrote more carefully

@Laurel: From the WaPo article: A full methodology can be found in their paper “Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models.”

@Oddthinking: The problem with both you and OP are that you are unable to think quantitatively and distinguish between orders of magnitude. To you and OP, "10 times more", "100 times more", "1000 times more", and "10000 times more" are all just "much more" and need not be distinguished.

@user182601 Did you find the information in there? All I can find about GPT-4 is "Due to the limited public data available for GPT-4, we now present a case study to estimate GPT-3's operational water consumption footprint". The estimate WaPo gives for 4 is nowhere to be found in the paper, at least not that I can find.

@Laurel: Water and electricity costs were calculated by Ren for ChatGPT-4 at an average American data center. I interpret this to mean that WaPo asked Ren separately to use their paper's methodology to compute the figures for a 100-word email

In financial terms, I expect enough meat to feed two people to cost on the order of 5-10,000 more than an email (e.g. $10 versus a milli-cent). That is a very weak proxy for the amount of physical resources, but it means the figure you quote doesn't surprise me at all.

An interesting comparison would be how much a regular email costs. That something grown over years takes much more resources than something generated in (sub)seconds should not be surprising. FWIW, that is a factor at the order of 10000000 to 100000000, just to think quantitatively.

The comparison here is rather apples to oranges. The equivalent of generating one email would rather be cutting a piece of beef, and the equivalent of growing beef would be building the data centre, producing its hardware and training the LLM.

We don't really have a choice about spending water and energy on food production if we want to survive, though there are certainly things that use less water than beef or rice (and many people advocate switching to them). By contrast, by using a much smaller amount on top of the energy and water that we are already using each day—the energy from that pound of rice, if you will—we can write hundreds of emails. Including the needs of essential goods in a comparison like this does not seem meaningful, because we cannot stop eating to produce more LLM-generated emails while maintaining costs.

Actual sincere advice to everyone in this thread. If you care this much about this little water, go find a water conservation org and figure out how to have actual impact. Humans in developed countries average 1 liter of water every 4 minutes just through daily activities and consumption, so our time is equivalent to water. If you spend 4 minutes writing an email, you've already used twice as much water as ChatGPT would have used. So use ChatGPT to write your email, and then with the 3 minutes you saved, go check your garden hose for leaks.

@TheGuywithTheHat - That doesn't seem like sincere advice. It seems more like an argument meant to dismiss the environmental concerns. And not one that is particularly sound, either: most people probably have enough spare time to check their garden hose for leaks regardless of what they do elsewhere, assuming that they even have one, so they can avoid using energy-and-water-inefficient tools and check the hose if they want. Even worse, taking your numbers as correct ("twice as much"), they could be framed as "GPT will increase your water use by 50% for every unit time that you use it."

@Obie2.0 If we go by these numbers (ChatGPT water use from top answer an human daily use from a quick google search), ChatGPT only has to save you 2 minutes per use for it to be a net water gain, assuming you go do something else productive with those 2 minutes you saved. If you truly can't think of anything better to do with your time than save 1 bottle of water by manually writing an email that could have been written by an LLM, then go for it, but just be aware that that amount of water is 0.1% of your daily usage. BTW this discussion has wasted water equivalent to ~50 uses of ChatGPT.

@TheGuywithTheHat - I don't think you understand what I am saying. You are not using substantially more water by writing these comments. You do not have 50 garden hoses to go fix, or if you do, you are an outlier. You seem to be under the impression that there is some activity you could be doing right now that would use substantially less water, or save you substantially, which would be what you would need for your argument to be correct. But if you are like most people, there isn't. You'll take a bath whether or not you write this comment, you'll turn on your air conditioning regardless.

And GPT models won't stop you from doing that. Now, if you want to go fix your garden hose with the free time that you have (that you have it is obvious), do that, but don't give LLMs credit for that choice.

@Obie2.0 I could install a shower, toilet, or faucet that uses less water. I could check all my pipes for leaks. I could find local food sources that use less water in packaging and transport. I could fix a broken item in my house rather than buying a new one. I could research which foods use less water to produce. I could replace my garden plants with ones that use less water. I could lobby my local gov to put restrictions on yard watering. I could donate to a national org that lobbies at the federal level. I could write comments on the internet telling others what they can do to save water.

AI water usage is not significant relative to many other industries. Worry burnout takes attention away from things that matter.

@Oddthinking: (1) We're comparing water, not dollars (in value? cost?) or "resources". Your arguments depend on this flawed reasoning: If it costs $30 to fill my 10,000L swimming pool while a back massage costs $60, then we'd expect the back massage to use up about 20,000L of water.

@Oddthinking: (2) Where did your "$10" and "milli-cent" figures come from? For the former, why not $100 or $1000? (I'm sure you can find meals for two with a pound of beef that cost $1000.) // For the latter, why not a millionth of a cent? Or a billionth? Or a zillionth or gazillionth? // As usual with you and others who are not accustomed to thinking quantitatively and making some effort at precision ("lots of" (OP), "much more" (you), "huge" (End Antisemitic Hate above)), numbers and orders of magnitude don't matter. You can just make up any number ("milli-cent") to suit your argument. ...

... If instead of lazily inventing numbers to suit your argument, you simply took a few seconds to google, you'd find that a pound of rice sold to retail consumers is $1 or less (example), and not $10 (the number you invented). At wholesale prices it'd be even cheaper. Numbers matter, orders of magnitude matter. They are facts that can often be discovered. Not inventions to suit your argument.

I have two apologies:

The first is the delay in responding. I have been travelling, and the nature of this discussion made it not a priority.

The fundamental problem with this answer that several people have expressed hasn't been addressed. I will try to explain again; hopefully it will be clearer.

One reason that would be a bad answer is that it is a false equivalence. Hitler et al didn't personally kill that many by hand. Likewise, comparing resource-intensive purified drinking water used by data centres to untreated rain water/dam water/river water used in farming is unreasonable. [Note: I haven't heard of data centres using river water, but I could be ignorant] The original claim is about "bottles of water" which suggests clean drinking water.

The other reason is it is comparing the outliers that we already acknowledge as bad. Just because one act isn't as morally bad as a holocaust doesn't make it morally good. Just because an AI query doesn't use as much resources as meat - a product that people many people boycott because of its resource usage - doesn't mean it doesn't use a lot more resources than people might expect.

This is what I referred to as "bickering about the definition of lots". It didn't focus on how much water is used by AI, and instead referenced fairly weak evidence about "meat" and "rice", which have wide variances in their water use.

@user182601: In defending yourself, you have been rather disparaging of others. Please keep an eye on that. Not only does it hurt our community, but it undermined your arguments.

> We're comparing water, not dollars (in value? cost?) or "resources".

Yes, absolutely. I agree completely with the commenter above who said the financial terms "is a very weak proxy for the amount of physical resources". [Hint: That was me.]

Nonetheless, if your attempt was to say "A pound of meat uses four orders of magnitude more water than an AI query. This shocking comparison is evidence that AI doesn't use lots of water." my point still stands. It isn't shocking because meat farming is famously resource-hungry.

> Where did your "$10" and "milli-cent" figures come from? For the former, why not $100 or $1000? (I'm sure you can find meals for two with a pound of beef that cost $1000.)

My second apology is that I misunderstood your email reference. I misunderstood, and thought you were talking about the cost of propagating an already written email, not using AI to generate one.

I tried to estimate, based on my understanding of IT costs, what the right order of magnitude was. (Email is often free to the consumer, but I tried to estimate what an ISP would charge, per day, to provide an SMTP server, or similar, to deliver 100,000 marketing newsletters a day, and ended up with a microcent estimate.)

I did not intend to estimate the cost to AI-generate 100 words of text, because that is very much the crux of the issue and deserves solid references. I didn't mean to beg the question. I am sorry for the confusion that caused.

I reject your argument that this is all about me being unaccustomed to thinking quantitively.

I reject your argument that this is about my laziness. I remind you to be careful of ad hominem attacks.

I reject your strawman about the cost of rice, because I was talking meat as an example.