Log in
I Wish I Could Stop Writing About AI
Recently, a bunch of folks on the fediverse shared what sounded like an interesting post, claiming that the environmental impacts of generative AI such as LLMs is overstated, especially when considering ChatGPT-style chatbots.
Andy Masley: Why using ChatGPT is not bad for the environment - a cheat sheet
After all, I've frequently argued that, at least in its modern incarnation, AI is a climate-destroying disaster. If I'm wrong in that, or even just possibly wrong, I'm obligated to at least do the due diligence to make sure I have my facts in order. Notably, that's not the same as "we should listen to both sides," fuck that noise. It's the basic admission that empiricism and rationality aren't foolproof, and it's possible to erroneously believe factual claims that wind up not being true.
Put differently, if someone claims that trans folks shouldn't exist, I don't owe it to anyone to hear out the details on that — I'm fully justified in telling them to fuck off right from go. On the other hand, if someone says "hey, trickle-down economics actually works, and here's some concrete data that proves it," I'll still almost certainly be justified in telling them to fuck off, but due diligence at least demands I understand where they're wrong.
In practice, I have limitations on my time, emotional energy, and just sheer willingness to engage with the same arguments over and over again. That is perfectly rational, I will posit, as a policy, even if it doesn't yield perfectly rational fact claims. So, in practice, I don't listen to the nerdier kinds of fascists claiming again and again to have proven that Reagan didn't do anything wrong. I'm not missing much.
With this post, though, arguments about generative AI and energy usage are about 35 years less mature, so I found myself clicking through to read the post. As one might expect from the title of my own post, and the fact that I've not done an about-face in public, admitting how wrong I was all this time, the post was more fact-shaped than factual. As a result, instead of writing what I'd like to write this afternoon, I find myself deep within my Someone Is Wrong on the Internet mode. Alas.
Let's Get This Out of the Way
Before I get into the nitty gritty of Masley's article, let me start by laying out a few basics about generative AI first. It's helpful in understanding what factual claims and objections are even at play here, and which are not under further contention at the moment. Indeed, as Masley himself says:
> This post is not about the broader climate impacts of AI beyond chatbots, or about whether AI is bad for other reasons (copyright, hallucinations, job loss, risks from advanced AI, etc.).
First and foremost, the term "AI" is itself rather confusing, and purposefully so! AI as a term goes back much, much further than the current hype cycle, and has expanded and contracted over the intervening years. Different subfields have adopted the term in different ways, too, confusing the issue further. For instance, game developers have historically used the term "AI" to refer to the set of programmed behaviors used to create the illusion that in-game characters are real in some limited sense. That is, from the perspective of video games, AI is almost a theatrical term.
It's of immense profit for the purveyors of generative AI products such as ChatGPT, Microsoft's various Copilots, Google Gemini, etc. to build on the long history of AI as a term to lend the latest round of bullshit some credibility. It's kind of like how whenever you point out that the police should get less funding due to all the violence and corruption, people point to police directing traffic. Yeah, that's good, but it didn't take having someone with a gun and a nearly unlimited license to kill showing up in the middle of an intersection. Indeed, quoting Masley again:
> The services using 97-99% of AI’s energy budget are (roughly in order): > > - Recommender Systems - Content recommendation engines and personalization models used by streaming platforms, e-commerce sites, social media feeds, and online advertising networks.
Recommender systems have existed for much longer than generative AI in its present form. The infamous Netflix Prize, designed to improve the performance of recommendation systems at the time, dates all the way back to 2006. Conflating generative AI and restricted Boltzmann machines augmented with decision trees is misleading at best.
Koren 2009: The BellKor Solution to the Netflix Grand Prize
To be clear, then, when I use the term "AI" in this post, I intend it in a narrower sense than Masley's use of the term, but closely in line with his focus on ChatGPT and similar chatbots that offer an interface to large language models. That itself is even a little vague, but such is the difficulty of trying to discuss the ever-shifting marketing claims made by giant tech companies.
With the term established, it's also critical to be clear about the dangers presented by AI (again, now in the narrow chatbot-like LLM sense, rather than the broader historical sense). AI products are trained on the uncompensated labor of millions, are promoted and used to further devalue and deskill jobs, and perhaps most chillingly, are intrinsically connected to the global rise of fascism. As I've argued before:
> I used to see the AI bubble and trans rights as distinct issues. I no longer do. The fascist movement in tech has truly metastasized, as evidenced by Elon Musk's personal coup, his endless supply of techbro supporters, tech companies' eagerness to axe DEI programs once Trump gave them an excuse, erasure of queer lives from tech products, etc. > > To the extent that AI marketing is an attempt to enclose and commodify culture, and thus to concentrate political power, I see it as a kind of fascism.
@xgranade@wandering.shop: "I used to see..."
My argument there is an echo of and has been echoed by others; I am far from alone in drawing a connection between eugenics, fascism, and AI.
> The AI projects currently mid-hype are being developed and sold by billionaires and VCs with companies explicitly pursuing surveillance, exploitation, and weaponry. They fired their ethics teams at the start of the cycle, and diverted our attention to a long-term sci-fi narrative about the coming age of machines – a “General Intelligence” that will soon “surpasses” human ability.
Miriam Eric Suzanne: Tech Continues to be Political
> I mean, every part of this is really upsetting to me, and I think that this notion of post humanity, which often goes along with AI, does embed within it some really troubling ideas about like the innateness of intelligence, whereas I think intelligence is in large part, a learned skill and it's a product of early childhood education and the kind of habits that you cultivate throughout your life of questioning things. But, you know, again, I'm just gonna restate, we don't know what intelligence is. We don't know how much of it is genetic or what kind of factors shape it. We can't even define it. And yeah, it's just… This notion that we can become super beings inevitably just goes to some really dark, really racist places.
Our Opinions are Correct, Episode 125: Silicon Valley vs. Science Fiction: ChatGPT
With that in mind, I want to revisit Masley's introduction before moving on to his factual claims:
> This post is not about the broader climate impacts of AI beyond chatbots, or about whether AI is bad for other reasons (copyright, hallucinations, job loss, risks from advanced AI, etc.).
That phrase, "risks from advanced AI," should raise some eyebrows given all of the above. Indeed, Masley links to an 80,000 Hours post about "AI catastrophe" as a citation for said "risks."
80,000 Hours: Preventing an AI-related catastrophe
For the unfamiliar, 80,000 Hours is a nonprofit dedicated to advancing the cause of Effective Altruism (EA), a philosophy sometimes derided as an "ethics of the rich."
PhilosophyTube: The Rich Have Their Own Ethics
80,000 Hours was indeed founded by none other than William MacAskill, the founder of the EA movement; it's no secret that I do not think very highly of him nor his apologia for the likes of Sam Bankman-Fried, one of his biggest supporters. But that's somewhat beyond the point here. Rather, the link between EA and a related movement known as longtermism is critical to understanding why including a link to 80,000 Hours is as suspect. While EA teaches that the rich should be as rich as possible, as the rich are the best arbiters of how to allocate societal resources for human flourishing, longtermism teaches that the foibles that we may suffer now are nothing compared to the risks of not developing general AI in the far, far future. Reading the post cited by Masley, then, we see that much of the risk highlighted by 80,000 hours consists of AI not being "aligned," a framing deeply rooted in EA and longtermism.
> When we say we’re concerned about existential catastrophes, we’re not just concerned about risks of extinction. This is because the source of our concern is rooted in longtermism: the idea that the lives of all future generations matter, and so it’s extremely important to protect their interests. > > This means that any event that could prevent all future generations from living lives full of whatever you think makes life valuable (whether that’s happiness, justice, beauty, or general flourishing) counts as an existential catastrophe.
As an aside, it's notable what risks 80,000 Hours does not include in the risks presented by AI. In their page on longtermism, they note that while climate change is very serious, it's not likely an extinction-level risk by their accounting:
> Climate change, for example, could potentially cause a devastating trajectory shift. Even if we believe it probably won’t lead to humanity’s extinction, extreme climate change could radically reshape civilisation for the worse, possibly curtailing our viable opportunities to thrive over the long term.
By comparison with supposedly existential risks such as AI becoming sentient in ways that don't benefit capital labor, responding to the climate emergency is more of a nice-to-have. I don't agree with those priorities, to say the least.
Notably, all of the above stands whether or not ChatGPT's environmental impacts are as benign as Masley claims, a point that he acknowledges in his very introduction, even if he cites the philosophical basis for modern tech-funded fascism in doing so (that the article is posted to Substack does not escape me). What, then, do we learn if Masley's claims hold? Not nothing, to be sure. As I mentioned at the get-go, acting in concert with the facts is pretty damned important. At the same time, our course of action should be clear regardless of the climate impacts of ChatGPT et al.
Joel Pett: What if it's a big hoax and we create a better world for nothing?
I argue that we shouldn't be tolerant of sloppy factual claims, let alone lies and disinformation, but we also need to keep perspective: it's worth opposing fascists even if they don't pollute that much, and it's worth protecting labor even if the externalities of doing so are fairly negligible. That is, I'll warrant, a somewhat subtle and nuanced position, but hey. This is my blog, so I get to have opinions that take more than a sentence or two to express!
How Much Energy Does a Prompt Cost?
Much of Masley's claims derive from a single central claim: that the cost of a single prompt to a typical ChatGPT-like chatbot is upper-bounded by 3 Watt-hours (3,600 Joules), or about the full capacity of a AA-sized Li-ion battery, and is likely to be far less than that overestimate.
I choose that comparison to be somewhat inflammatory, of course. Masley chooses other comparisons to illustrate his point, including "running a microwave for 10 seconds." The daily life of a typical North American includes energy references at massively different scales, offering a lot of opportunity to choose more or less innocuous-sounding comparisons. A typical plugin hybrid car may have a capacity of about 40 kWh, or 144,000,000 Joules; that's roughly 40,000 ChatGPT prompts using Masley's claimed figure, so who cares, right?
There's three problems with this analysis, though: we don't actually know how much energy a ChatGPT prompt costs, that figure doesn't include the cost of collecting data for and training an AI model, and own its own, that analysis doesn't suggest any particular course of action.
For the first, Masley cites only Epoch AI, an AI industry research institute. Of particular note, Epoch is a group with a vested financial interest in reaching a particular set of conclusions, namely that AI is an industry that should see further investment. While that conflict alone doesn't invalidate their conclusions, taken together with the scarcity of corroborating data, it warrants some modicum of suspicion. Where do they get that 3 Watt-hour upper bound, then? Epoch proceeds in three steps: estimating the number of parameters that must be evaluated for a given chunk of output, estimating the number of expected chunks of output, and estimating the energy cost per parameter evaluation. This strategy necessarily involves making a significant number of assumptions, such as the duty cycle of GPUs in a data center, the ratio of average to peak dissipated power, and so forth. Epoch's assumptions may or may not be reasonable, but without more transparency into how ChatGPT and other chatbots are implemented, their assumptions are just that: assumptions.
Epoch, and by extension Masley, claim that 0.3 Watt-hours (360 Joules) per prompt would be a more accurate estimate, contrasted with those estimates obtained by de Vries in 2023.
de Vries: The growing energy footprint of artificial intelligence
Both the Epoch and de Vries estimates largely exclude energy costs other than the GPUs used in evaluating LLM outputs; a more comprehensive estimate would also include other costs such as CPU, storage, networking, and cooling. Regardless, arriving at a concrete estimate without access to data currently withheld by AI companies is difficult at best. Indeed, as de Vries has noted:
> This energy calculation felt like “grasping at straws”, de Vries says, because he had to rely on third-party estimates that he could not replicate. And his numbers quickly became obsolete. The number of servers required for an AI-integrated Google search is likely to be lower now, because today’s AI models can match the accuracy of 2023 models at a fraction of the computational cost, as US energy-analyst firm SemiAnalysis (whose estimates de Vries had relied on) wrote in an e-mail to Nature.
Nature: How much energy will AI really consume? The good, the bad and the unknown
Masley by and large shrugs off this problem, and assumes that the 3 Watt-hour per prompt figure is good enough:
> So I disagree that this is vibes and guesswork. It’s very uncertain! But people more knowledgeable than me have tried their best to put a number on the energy and water involved, so it’s more than a random shot in the dark. I’ve tried to defer to where all their guesses are. Almost all conversations about individual climate impacts from using ChatGPT seem to assume the same numbers are correct, so this is what’s being debated. We could all be wrong, but it seems just as likely that ChatGPT actually uses less energy than that it uses more. Given that we’re uncertain, and a lot of people are still making strong claims that ChatGPT is terrible for the environment, I think I’m perfectly within reason to write about how the numbers we have strongly imply that they’re wrong.
And now we're very firmly into where Masley's article becomes more fact-shaped than factual, more opinion than objective truth. That's fine, insofar as it goes, but he makes an implicit appeal to the authority that would come along with an objective analysis. In particular, if we're worried about whether or not something is, as Masley puts it, "terrible for the environment," then his own stated philosophies of effective altruism and longtermism as stated on 80,000 Hours would seem to imply that we should be cautious in the absence of objective data. At the very least, a more reasoned analysis would likely include a call for AI vendors to be much, much more transparent about the potential energy costs of their products.
I don't know how much energy a given prompt takes, and nor does Masley. That should be a point of concern, not something to blithely dismiss. Given the massive draw of modern data centers, and how much that draw has corresponded with increases in AI adoption, we at least have a rational basis to be suspicious of the rather rosy AA-battery-per-prompt figures, to say nothing of Masley's much more optimistic 0.3 Watt-hours per prompt assumption.
Recall from above that Masley claims to address that suspicion, however:
> The mistake they’re making is simple: ChatGPT and other AI chatbots are extremely, extremely small parts of AI’s energy demand. Even if everyone stopped using all AI chatbots, AI’s energy demand wouldn’t change in a noticeable way at all. The data implies that at most all chatbots are only using 1-3% of the energy used on AI.
I'll straight up say it: I have no idea where he's getting the 1 to 3% figure. My best guess is that this claim comes from the back of the envelope calculation in another of his posts. If so, that presents a problem for his analysis in that it assumes the very answer he's looking for!
More concerning, the 1% to 3% claim doesn't fit with energy usage doubling over a five-year timespan.
IEA: AI is set to drive surging electricity demand...
Here again, the secrecy of the AI industry makes it difficult to reach firm and objective conclusions. What we do know is that the companies who make and sell AI products have also been demanding startling amounts of energy, and claim that they need to do so to cover AI usage.
What Should We Do About It?
On the basis of his assumptions about the energy required for ChatGPT prompts and the volume of data center usage corresponding to chatbot prompting, Masley offers a very strong conclusion right from start:
> By being vegan, I have as much climate impact as not prompting ChatGPT 400,000 times each year (the water impact is even bigger). I don’t think I’m going to come close to prompting ChatGPT 400,000 times in my life, so each year I effectively stop more than a person’s entire lifetime of ChatGPT searches with a single lifestyle change. If I choose not to take a flight to Europe, I save 3,500,000 ChatGPT searches. this is like stopping more than 7 people from searching ChatGPT for their entire lives. Preventing ChatGPT searches is a hopelessly useless lever for the climate movement to try to pull. We have so many tools at our disposal to make the climate better. Why make everyone feel guilt over something that won’t have any impact?
This is, to be blunt, patent bullshit. The question that Masley sets out to answer is emphatically not whether or not to eat meat, but whether or not using ChatGPT et al. are likely to present a risk to the climate emergency. The existence and scope of other climate risks is absolutely immaterial to that question. Masley's insistence on making these kinds of distracting comparisons grows more absurd as his article grows in length:
> If everyone in the world stopped using ChatGPT, this would save around 3GWh per day. If everyone in the world who owns a microwave committed to using their microwaves for 10 fewer seconds every day, this would also save around 3GWh per day.
Is Masley taking action to encourage people to turn their microwaves on for 10 seconds less each day? He's absolutely arguing that people should use ChatGPT, even if only in the indirect and implicit form of downplaying the climate risks that doing so presents. If that doesn't also come with encouraging people to use the microwave less, then by his own argument, he's contributed to the waste of 3 gigawatt hours per day (125 megawatts, or about the capacity of a coal power plant).
Put this way, the absurdity becomes clear: these are incomparable activities, and whether or not Masley argues for more efficient microwave usage has no bearing on the efficiency and costs of ChatGPT. Here again, Masley seems to preemptively acknowledge the criticism, that committing to category errors in offering comparisons does not lead to any particular insight into how or if we should use ChatGPT:
> “Whataboutism” is a bad rhetorical trick where instead of responding directly to an accusation or criticism, you just launch a different accusation or criticism at someone else to deflect. Kids do this a lot.
Rather than actually understanding why "whataboutism", as he puts it, is bad, Masley simply redefines away his attempts at slight of hand:
> Under this revised definition, it’s whataboutism to say “eating meat isn’t bad because people drive,” but it’s not whataboutism to say “Google isn’t bad because its emissions are so drastically low compared to everything else we do,” and it’s not whataboutism to say the same about ChatGPT.
That is, if we take his premise, comparisons between different modes of energy consumption are always relevant, even making such comparisons offers no actionable insight.
Looking at his claims again in this light, let's revisit his list of comparisons for other activities that use approximately 3 Watt-hours worth of energy:
> - Leave a single incandescent light bulb on for 3 minutes. > - Leave a wireless router on for 30 minutes. > - Play a gaming console for 1 minute. > - Run a vacuum cleaner for 10 seconds. > - Run a microwave for 10 seconds > - Run a toaster for 8 seconds > - Brew coffee for 10 seconds
There's a massive difference between each of these points of comparison and a ChatGPT prompt. I like having light in my house, it's pretty useful. Having access to the internet is pretty wonderful, too. Playing video games is actually quite fun, and having a clean house is quite nice. It's great to have coffee and hot food on demand.
By contrast, ChatGPT does.... what? Masley is at our rescue with an answer!
> ChatGPT could write this post using less energy than it takes you to read it.
Ah. It could generate more disingenuous nonsense for the rest of us to sort through. Wonderful.
The Dismal View
Look at Masley's comparison again, and it becomes clear that it's even more dismal than it initially seems.
So fucking what if ChatGPT could write a blog post using less energy than a human? You don't give humans energy in the form of light, food, water, and heat so that they produce blog posts hyping your favorite tech nonsense, you give humans the energy to live because it's good to provide for fellow humans! Masley's throwaway joke there belies a much more bleak and dismal view still, that it is good and appropriate to view humans themselves as climate costs. Not our habits, our technologies, our cultures, but our mere existances. That kind of thinking is, without hyperbole, exactly in line with the kinds of Malthusian views of population dynamics that have motivated eugenics and fascist movements for a century.
> We will suppose the means of subsistence in any country just equal to the easy support of its inhabitants. The constant effort towards population...increases the number of people before the means of subsistence are increased. The food therefore which before supported seven millions, must now be divided among seven millions and a half or eight millions.
Masley's invocation of eugenicist thought here is far from unique, however.
> Our findings reveal that AI systems emit between 130 and 1500 times less CO2e per page of text generated compared to human writers, while AI illustration systems emit between 310 and 2900 times less CO2e per image than their human counterparts.
What a dismal, dreary view of art, culture, and human joy! I do not write to produce capitalist gain, I write because I am moved to, because it is one of the most human things in the world to tell stories— it's right up there with feeding, fighting, and fucking as far as imperatives go. As awful as that mindset is, though, it is deeply interwoven into the strange alliance between eugenicist movements, AI evangelism, longtermism, and effective altruism, as Our Opinions Are Correct called out elsewhere in episode 125. It's worth listening to the episode in its entirety, but if you'll forgive a longer excerpt, Annalee Newitz and Charlie Jane Anders deal with that view as epoused by Nick Bolstrom, author of Superintelligence:
> Annalee: So, keeping that in mind. Charlie Jane, I'm gonna have you read this quote from Superintelligence about how we will deal with super intelligent AI workers. > > Charlie Jane: Okay. “A salient initial question is whether these working machine minds are owned as capital (slaves) or are hired as free wage laborers. On closer inspection, however, it becomes doubtful that anything really hinges on this issue. There are two reasons for this. First, if a free worker in a Malthusian state gets paid a subsistence level wage, he will have no disposable income left over after he has paid for food and other necessities. If the worker is instead a slave, his owner will pay for his maintenance, and again, he will have no disposable income. In either case, the worker gets the necessities and nothing more. > > “Second, suppose that the free laborer were somehow in a position to command an above subsistence level income, perhaps because of favorable regulation. How will he spend the surplus? Investors would find it most profitable to create workers who would be “voluntary slaves” who would willingly work for subsistence level wages. > > “Investors may create such workers by copying those workers who are compliant. With appropriate selection and perhaps some modification to the code, investors might be able to create workers who not only prefer to volunteer their labor, but would also choose to donate back to their owners any surplus income they might happen to receive. > > “Giving money to the worker would then be, but a roundabout way of giving money to the owner or employer, even if the worker were a free agent with full legal rights.” > > Oh gosh, that is so dystopian. First of all, the notion that the only difference between a slave and a free work worker is how much resources you receive. Like if you're a free worker, you might get something above your subsistence needs. There's nothing about the actual nature of slavery, which is that you can't change jobs and you can't have a life or determine your own destiny. > > That's incredibly dark and weird in the notion that like, well, workers might be able to get paid more than subsistence level, but then we just turn around and make workers who are happy to work for free. > > And I'm like, that is some real rhetorical slippage. Like he's just like, oh, but then blah. Then we're just gonna magically create slaves anyway. And it's just, I don't even understand this paragraph. I've read it like three or four times and it just baffles me more and more each time.
Wrapping Up
Masley's post was shared fairly widely, and by some quite influential folks in tech, such that I think it was worth a few thousand words to critically examine his central claim: that we should not refuse to use ChatGPT and similar AI products on an environmental basis alone. Trying to understand that claim necessitates a bit of a detour into what AI is, how it intersects with eugenicist and fascist movements in tech, and about the climate emergency in general. Masley's central claim, however, rests entirely on guesswork (however reasonable!) and, to borrow his own term, "whataboutism."
His argument, then, is much better served by advocating for AI vendors to provide the transparency needed to actually evaluate the climate impact of LLM usage. I would personally be far more sympathetic to his post had he taken that tack, rather than downplaying the very real concerns raised with LLM products.
As it stands, though, I submit that Masley's arguments are best understood as fact-shaped apologia rather than a serious contribution to discourse. Perhaps even more unfortunately, this dismissal comes along with some truly depressing statements about the human condition; statements that I find downright appalling as someone who spends much of their day learning how to write and communicate with fellow humans.
In light of all the above, I'm happy to shut the proverbial book on this post, and be content continuing to oppose the proliferation of AI products.
Support me!
This is a longer post than my normal, by a fair bit, and took me several hours to write and research. If that is valuable to you, please consider supporting that labor. Thank you so much!
