> AI isn’t especially environmentally unfriendly

I think the actual answer is more nuanced and less positive. Although I appreciste how many citations your comment has!

I'd point to just oe, which is a really good article MIT's technology review published about exactly this issue[0].

I'd make two overall points firstly to:

> when you have the opportunity to use human labour or AI, AI is almost certainly the greener option as well.

I think that this is never the trade off, AI normally generates marketing copy for someone in marketing, not by itself, and even when if it does everything itself, the marketing person might stop being employed but certainly doesn't stop existing and producing co2.

My point is, AI electricity usage is almost exclusively new usage, not replacing something else.

And secondly on Simon Wilison / Sam Altman's argument that:

> Assuming that higher end, a ChatGPT prompt by Sam Altman's estimate uses: > > 0.34 Wh / (240 Wh / 3600 seconds) = 5.1 seconds of Netflix > > Or double that, 10.2 seconds, if you take the lower end of the Netflix estimate instead.

This may well be true for prompts, but misses out the energy intensive training process. Which we can't do if we actually want to know the full emmisions impact. Especially in an environment when new models are being trained all the time.

On a more positive note, I think Ecosia's article makes a good point that AI requires electricity, not pollution. It's a really bad piece of timing that AI has taken off initially in the US at a time when the political climate is trying to steer energy away from safer more sustainable sources, and towards more dangerous, polluting ones. But that isn't an environment thay has to continue, and Chinese AI work in the last year has also done a good job of demonstrating that AI trainibg energy use can be a lot kess than previously assumed.

[0] https://www.technologyreview.com/2025/05/20/1116327/ai-energ...

I think this article is a good response to the MIT article: https://andymasley.substack.com/p/reactions-to-mit-technolog...

> AI normally generates marketing copy for someone in marketing, not by itself, and even when if it does everything itself, the marketing person might stop being employed but certainly doesn't stop existing and producing co2.

Sure, but it does it a lot quicker than they can, which means they spend more of their time on other things. You’re getting more work done on average for the carbon you are “spending”.

Also, even when ignoring the carbon cost of the human, just the difference in energy use from their computer equipment in terms of time spent on the task outstrips AI energy use.

> This may well be true for prompts, but misses out the energy intensive training process.

If you are trying to account for the fully embodied cost including production, then I think things tilt even more in favour of AI being environmentally-friendly. Do you think producing a Netflix show is carbon-neutral? I have no idea what the carbon cost of producing, e.g. Stranger Things is, but I’m guessing it vastly outweighs the training costs of an LLM.

There's probably a decent chance that training an LLM produces more carbon than producing stranger things

I’ll readily admit that I don’t know the first thing about television production, but that doesn’t seem plausible to me. Moving lots of physical objects around takes far, far, far more work than shuffling bits, and a large proportion of that can’t come from sustainable energy sources. Think about things like flying the cast to shoot on location in Lithuania, for instance. Powering and cooling servers isn’t in the same ballpark.

Not for nothing but the vast majority of people doing the kind of work that’s done on TV & Film just-so-happen to be geographically co-located for some reason.

It’s possibly worth noting that both activities require humans and even fully operational end-to-end supply chains of rare-earth minerals and semiconductor fabrication. Among many, many other things involved.

I just don’t think we can freely discount that it takes heavy industrial equipment and people and transport vehicles to move and process the raw materials to make LLM/AI tech possible, and that the … excitement has driven those activities to precipitous heights. And then of course transporting refined materials, fabricating end products, transporting those, building and deploying new machines in new data centers around the world, massively increasing global energy demand and spiking it way beyond household use in locales where these new data centers are deployed. And so on and so forth.

I suspect that we will find out someday that maybe LLMs really are more efficient, possibly even somehow “carbon negative” if you amortize it across a long enough timespan—but also that the data will show, for this window of time, that it was egregiously bad across a full spectrum of metrics.

> there's probably a chance

It's this completely unfounded barrage of making shit up about energy consumption without any tether to reality that makes the whole thing with complaining about energy use seem just like a competition on who makes up the most ridiculous most hand-wringing analogy.

Glad to see someone refute the AI water argument, I'm sick of that one. But I do not see how the displacement argument fits. Maybe you can elaborate but I don't see how we can compare AI usage to watching Netflix for any length of time. I can't see a situation where someone would substitute watching stranger things for asking chatGPT questions?

The writing and illustrating activities use less energy, but the people out there using AI to generate ten novels and covers and fire them into the kindle store would not have written ten novels, so this is not displacement either

> I can't see a situation where someone would substitute watching stranger things for asking chatGPT questions?

Really? A bored kid can’t play around with ChatGPT instead of watching Netflix?

You can substitute any other activity if you like. Netflix was just an example.

Well the point was that chatgpt doesn't use so much energy. I thought we were comparing similar activities. If we're just going to list things a bored kid can do instead of watch Netflix we'll be here all day, and most of them use less electricity than chatGPT.

How many tens of thousands more pages of text and image are churned out per human created page though?

Pages that would never be created were the stochastic parrot to be turned off and never squawk.

It'd save a lot of energy, water, carbon emissions to just let the already existing humans just get on with the churn.

> How many tens of thousands more pages of text and image are churned out per human created page though?

I don’t know, how many?

> It'd save a lot of energy, water, carbon emissions to just let the already existing humans just get on with the churn.

How much, and how do you know that?

Please read the source I linked to; your reply doesn’t make any sense in that context.

I did, you clearly didn't.

> For instance, the emission footprint of a US resident is approximately 15 metric tons CO2e per year22, which translates to roughly 1.7 kg CO2e per hour

Those 15,000kg of CO2e are emitted regardless of that that person does.

The article also makes assumptions about laptops that are false.

>Assuming an average power consumption of 75 W for a typical laptop computer.

Laptops draw closer to 10W than 75W, (peak power is closer to 75W but almost not laptops can dissipate 75W continually).

The article is clearly written by someone with an axe to grind, not someone who is interested in understanding the cost of LLM's/AI/etc.

It says that ignoring the human carbon use, just their computer use during the task far outweighs the AI energy use. So your response “are you planning on killing the human?” makes zero sense in that context. “They are wrong about the energy use of a laptop” makes more sense , but you didn’t say that until I pushed you to actually read it.

75W is not outlandish when you consider the artist will almost certainly have a large monitor plugged in, external accessories, some will be using a desktop, etc. And even taking the smaller figure, AI use is still smaller.

The human carbon use is still relevant. If they were not doing the writing, they could accomplish some other valuable tasks. Because they are spending it on things the AI can do, somebody else will have to do those things or they won’t get done at all.

75w is nuts actually. I measured my _desktop_ setup about 10 years ago including two monitors and idle was around 35w. It also doesn't make sense to include idle of all peripherals since you would be using them for chatgpt as well.

Why do you think idle is the relevant figure here? They will be actively using the computer and its peripherals.