Why Using ChatGPT Is Not Bad for the Environment
At his Substack newsletter The Weird Turn Pro, Andy Masley presents a comprehensive argument that concerns about ChatGPT’s environmental impact are misplaced:
The question this post is trying to answer is “Should I boycott ChatGPT or limit how much I use it for the sake of the climate?” and the answer is a resounding and conclusive “No.”
It’s not bad for the environment if you or any number of people use ChatGPT, Gemini, Claude, Grok, or other large language model (LLM) chatbots. You can use ChatGPT as much as you like without worrying that you’re doing any harm to the planet. Worrying about your personal use of ChatGPT is wasted time that you could spend on the serious problems of climate change instead.
I’ve received some questions and comments about the environmental impact of AI chatbots and imagebots, so I was pleased to find Masley’s detailed analysis. If you’re interested in the topic, the linked article is just the cheat sheet for a longer 9,000-word article on the topic, and he has also responded to criticisms of his posts.
His conclusion is that the amount of energy and water consumed by a ChatGPT prompt, while real when multiplied across 1 billion prompts per day, isn’t worth worrying about compared to other activities. One ChatGPT prompt is roughly equivalent to running a vacuum cleaner for 10 seconds or a laptop for 3 minutes—or possibly even ten times less. Masley calculates that, on a daily basis, the average American uses enough energy for 10,000 ChatGPT prompts and consumes enough water for 24,000–61,000 prompts. When it comes to addressing climate change, we would be better served to focus on systemic change, or at least other personal lifestyle decisions.
While there is no requirement for anyone to use a chatbot or AI answer engine, if you are exploring their capabilities, you can do so without worrying that they will add to your existing climate footprint.
I think Masley’s Substack posts point out a lot of the challenges humans face in thinking about complex issues, especially when one holds a strong, emotionally-based opinion. I also think the hostility towards expertise that has emerged over the last decade or so in the US is a gigantic impediment to critical thinking.
For generative AI, I think a confluence of anti-technology views, anti-business views, and innumeracy lead to a lot of the “ChatGPT is destroying the environment” rhetoric. Further, the tendency of many to only look at single aspects of intricate systems, such as the manufacturing and use of glass and metal water bottles and Amazon’s logistics operations, can also lead to emotionally attractive but marginally effective calls to action.
Electron generation (electricity) DOES have an environmental impact no matter how it’s generated. So, conversing with an AI pollutes. Just like my reply.
Yes, but it’s a matter of proportion.
The 13W LED bulbs in my office use power, and less will be used if I turn them off when I step out for a few minutes. But my home’s HVAC system consumes orders of magnitude more power. It’s highly unlikely that the amount saved by avoiding lightbulbs altogether would even be noticeable.
Therefore, you should keep your A/C set at 85 in the summer and 65 in the winter. If you care, that is.
Funny enough: I’ve fed ChatGPT the article above and asked it about its energy consumption compared to the use of Google or Ecosia. And guess what ChatGPT told me
( The conclusion comes after the comparison ! )
TL;DR: chatgpt uses significantly more energy and water the “normal” search engines
Ecosia:
##
Conclusion
If you’re looking for quick facts or an overview of a topic:
Perhaps rerun your prompt to ask for comparisons with everyday activities to put the energy and water into context. 30 times nearly nothing is still nearly nothing.
Personally, since researching organizations’ offset programs is similar to assessing how non-profits use donations and takes a similar level of effort, I rarely use offsets as a single determining factor in use/don’t use decisions. For example, if I need to fly somewhere and Airline A and Airline B both offer offsets, I’m simply going to choose the flight that fits my schedule the best.
I was trying to figure out energy usage recently. I turned to ChatGPT.
My comparison:
I haven’t checked these numbers, but:
Okay. But asking ChatGPT about itself is inherently hazardous, because it’s simply answering based on its LLM training. If there’s been a plethora of articles, posts, and back-of-the-envelope calculations that it has trained on, it will answer based on the higher probability that the string of phrases it’s following are plausible.
Not that it’s wrong, but mentioning “Ecosia” in your prompt certainly seemed to produce a response that could have been written by Ecosia’s PR team.
Which makes sense. Do a web search for “Ecosia”. The overwhelming majority of hits are from Ecosia’s own web sites. So I would expect an LLM to generate text mostly based on the content of those sites.
ChatGPT, on the other hand, has been the subject of countless news and review articles, so there is a lot more than marketing material for an LLM to draw from.
Not entirely true now that ChatGPT searches the Web too. I find that most of my usage triggers Web searches automatically, though occasionally I notice that it didn’t base on a search, so I tell it to try again with one, causing the results to change noticeably with real-world information.
Tree planting is not a panacea. It is controversial.
I have my A/C set at 78F/25C in the summer and the heating at 68F/20C in the winter. I live in central AZ at 5,300 ft/1,615 meters elevation so I don’t get the triple digit Fahrenheit temperatures they get in the Valley, or the large amounts of snow they get up in Flagstaff.
AI, like everything else we rely on, uses energy—but it’s important to put that in perspective.
What often gets missed is that AI is already driving significant energy savings across sectors like transportation, logistics, manufacturing, and even energy production itself. From optimizing delivery routes and reducing fuel consumption to improving factory efficiency and helping balance power grids, AI is cutting waste in ways that scale.
There’s every reason to believe those savings will continue to accelerate—especially as AI systems become more efficient and begin optimizing their own operations.
The Economist recently published a couple of thoughtful articles on this topic, if you’re interested in digging deeper.
Would appreciate links. Thanks!
The language in your post is a bit too complex for the average reader to easily follow. So, I asked ChatGPT to rewrite it using simpler and more concise language—a common practice in the business world to improve clarity and accessibility.
“I think Masley’s Substack posts do a good job showing how hard it is for people to think clearly about complex topics—especially when they already have strong feelings about something. I also think that in the U.S., a growing distrust of experts over the past decade has made critical thinking even harder.”
When it comes to generative AI, I think a mix of anti-technology attitudes, distrust of big business, and poor understanding of numbers has fueled a lot of the “ChatGPT is ruining the planet” talk. Also, people often focus on just one part of a bigger system—like how glass or metal water bottles are made, or how Amazon delivers packages—which can lead to emotionally appealing but not very useful solutions.
I don’t know where chat bots end and AI in general begins. But there have been many reports of needing more energy generation to support growing AI use.
Adding power plants to meet the energy requirements of AI, implies that AI is a significant energy consumer.
The need for a lot more electricity (and therefore a more power plants) is vastly bigger than just AI. Some other huge examples:
Note that many of these transitions (especially home heating and vehicles) are being strongly encouraged (and sometimes mandated) by laws and government regulations.
When combined with the fact that states (again, often because of laws and regulations) are shutting down coal, oil, gas and nuclear power plants, it should not be the least bit surprising that there’s a massive energy crisis coming. And the only solution (which very few people in government want to support) is to build new power plants and expand the grid.
AI, while possibly significant, isn’t going to be the cause of or the solution to this problem.
I’m glad someone beat me to posting the MIT Technology Review article. It’s another good look at things, though they don’t seem to acknowledge that the industry actively wants to reduce power use as well. Or, rather, the industry actively wants the code to become more efficient because then they can spend less on power per prompt. My understanding from the API costs of models (how much it costs per prompt when you’re paying for each one) is that they have dropped radically over time.
Following the sentiment expressed in your post, here are two other ChatGPT rewrites that you may find more accessible:
Winston Churchill
It is my firm belief that Masley’s insightful posts on Substack illuminate the formidable challenges that humanity faces when confronted with complex issues, particularly when one clings tenaciously to strong, emotion-driven opinions. Furthermore, we must acknowledge the growing hostility towards expertise that has taken root in our society over the past decade—a hostility that stands as a colossal barrier to the exercise of critical thinking.
In the realm of generative AI, we find ourselves beset by a confluence of anti-technology sentiments, aversions to business, and a troubling lack of numerical literacy. This has given rise to the alarming rhetoric that “ChatGPT is wreaking havoc upon our environment.” Moreover, we observe a troubling tendency among many to focus solely on isolated aspects of intricate systems—be it the production and use of glass and metal water bottles or the complex logistics of Amazon. Such narrow perspectives may yield emotionally compelling arguments, yet they often result in actions that are, at best, marginally effective. Let us strive for a broader understanding, for it is only through comprehensive insight that we may navigate the challenges before us.
Quentin Tarantino
INT. COFFEE SHOP - DAY
The camera zooms in on a table where two friends, JACK and LUCY, sit sipping their coffee. JACK leans in, animated, as he talks about Masley’s Substack posts.
JACK
(leaning forward, intense)
You know what Masley’s been laying down on Substack? It’s like a spotlight on the mess we humans make when we tackle complex issues. I mean, when you’ve got a strong opinion fueled by emotion, it’s like trying to drive a car with no brakes. You just crash.
LUCY nods, intrigued, her eyes wide.
LUCY
Yeah, but what about the hostility? It’s like a plague.
JACK
Exactly! This hostility towards expertise that’s been brewing over the last decade in the U.S.? It’s a massive roadblock to clear thinking. Like, come on! We need experts, not just loud voices.
He takes a sip of his coffee, then leans back, crossing his arms.
JACK
And let’s talk about generative AI. It’s a cocktail of anti-tech vibes, anti-business rants, and a whole lot of people who can’t count. You hear the chatter? “ChatGPT is destroying the environment!” It’s like a bad movie line.
LUCY chuckles, shaking her head.
LUCY
Right? It’s all surface-level drama.
JACK
(gesturing wildly)
Exactly! People only see one piece of the puzzle. They’re fixated on glass and metal water bottles or Amazon’s logistics like it’s the whole story. But it’s not! It’s a tangled web, and they’re just pulling at one thread.
He leans in closer, eyes narrowing.
JACK
And those calls to action? They sound good, but they’re like a flashy car with no engine. Emotionally appealing but barely effective.
LUCY raises her coffee cup, a smirk on her face.
LUCY
To the tangled web, then.
They clink their cups together, the camera pulling back as they continue their animated discussion, the world around them fading into the background.
FADE OUT.
Some actual data from Sam Altman:
(Altman’s figures)
At a billion queries a day, that works out to 340 MW and
85,000 gallons of water per day.
Some context from a cursory look at search results:
Using water consumption data from the Commercial Buildings Energy Consumption Survey (CBECS), EIA estimates that the 46,000 [1] large commercial buildings (greater than 200,000 square feet) used about 359 billion gallons of water (980 million gallons per day) in 2012. This level represents an estimated 2.3% of the total public water supply in the United States [2]. On average, these buildings used 7.9 million gallons per building, 20 gallons per square foot, and 18,400 gallons per worker in 2012. On a daily basis, they used an average of 22,000 gallons per building, 55.6 gallons per thousand square feet, and 50.1 gallons per worker.
https://www.eia.gov/consumption/commercial/reports/2012/water/
On average each square meter of a hotel room uses around 0.55 kWh each day. So, the amount your hotel consumes will be based on the number and size of rooms. If you have a standard hotel room size (for example 91 M hotel with over 100 rooms could expect to be consuming around 50,000 kWh per day just for the hotel rooms.
Mistral AI have audited their water, etc, usage
https://bsky.app/profile/emollick.bsky.social/post/3luljwvstrs2d
Yes this is true.
There are other factors that cannot be accounted for, such as “competitive” behavior and organizations that do whatever they want and claim different (or opposite). This is becoming clear with the X (Grok) supercomputer complex in Memphis, TN, although there were many warnings over the last year as they raced to build their system and “beat” the industry at any cost.
An article from yesterday’s Washington Post (Gift Link[1]) puts into perspective the climate impact of AI vs other aspects of daily computing:
Moreover, the climate impact of AI is insignificant relative to other sources of digital emissions:
And digital emissions are insignificant relative to other personal choices:
Articles accessed via Gift Links are free to read without a subscription, but require a Washington Post account, which is also free. ↩︎
AI usage has nothing on pets.
The average dog (they eats lots of meat, especially beef) renders about 770 kg CO2 annual emissions. A large dog about 1500 kg. For comparison, the average US car used for the average US distance comes in at 4600 kg. If you’d ask some AI/LLM two dozen queries every single day that would amount to 34 kg. A cat BTW still clocks in at 310 kg (indoor, worse if outdoor).
I don’t know about where you guys live, but here in the Bay pets are sacred. Nobody would forgo their dog because of climate change even though we’re usually good at talking a big game about it. So I do wonder when AI keeps being brought up in the context of GHG emissions, if this is more about something else than how we actually curb climate change.
(And just in case you’re truly concerned about the GHG emissions of those 24 queries every single day, next year just drive 85 miles less. That’s merely a quarter mile a day walking instead of driving. Done.
Time to put the kill switch on the TV outlet!
Fine idea except that TVs are not progressing in that direction.
It’s unlikely folks will want to kill power to their “smart” TVs, if it then takes two plus minutes to turn it back on (booting Android and loading all the junkware etc.)
I’m having a hard time reconciling the “AI energy use is truly insignificant” with the claims by the big AI companies that they…
I think the incongruity is because building, training, and storing LLMs and generative AI’s is highly energy intensive but individual queries are less so.
Similarly, mining bitcoin is a huge energy suck but bitcoin
transactionstransaction queries are not.(Thanks to @Shamino for refining my comparison)
That’s true of a lot of environmental impacts – driving to the store to get milk is not a massive use of energy. 300 million people driving to the store is.
That’s a bit misleading.
Querying the Bitcoin blockchain to look up/verify transactions is lightweight.
But any transaction that puts another block on the chain (that is, every time bitcoins are transferred from one account to another), requires the mining process (including all of its overhead) in order to complete the transaction.
But then shouldn’t that energy use count? If the argument is “using AI isn’t very energy intensive so you shouldn’t worry about it,” then is it fair to say we’ll just ignore massive infrastructure costs that will have to be built out to support that use?
Is it okay to trade in your Prius for an Escalade with the emissions controls disabled because, after all, how much of a difference will one excessive driver and one particularly dirty vehicle make overall?
Just to clarify—I don’t want to debate whether LLMs/gen AI’s are “good” or “bad” based on power requirements—my intention was to address why companies including Meta and Microsoft can say they need new energy infrastructure while online commentators can also assert that usage of products such as ChatGPT and Perplexity is not as threatening to the environment as some people say.
It’s not just AI models. Just about everything everybody does today involves one or more cloud-based services. This means servers running in data centers, whether they’re running AI models or simple shell scripts or anything in between.
ChatGPT may consume a lot of power for the aggregate of its users, but so do the non-AI services that manage, for example, everybody’s smart lightswitches or doorbell cameras or video livestreams or on-line gaming, etc.