Introduction
As you have likely worked out on your own already, Artificial
Intelligence is not going away. It has gone from being a joke to a novelty to
the bogeyman to a tool that many of us use all the time. And yet, there are
still holdouts, perhaps you among them. In my workplace, it’s a mixed bag. As
recently as 2024 I was forbidden to use ChatGPT or other AI platforms at work;
now, my employer is wheedling,
exhorting, begging, and all but requiring my colleagues and me to adopt it. On
the family front, one of my daughters uses it a fair bit (sometimes frivolously),
the other not at all. My wife is wary of it.
So should you use
AI? I consider myself fairly well qualified to answer this. I have been
dabbling in AI for almost fourteen years; have devoted a fair amount of
research to kicking its tires; and now use it extensively both at work and at
home. I’ve blogged about it a bunch of times. I’m unbiased, since I don’t work for the AI Industrial
Complex, but I also don’t have a knee-jerk fear of technology.
I’ve blogged before (here)
about how we can use AI, describing two
fundamental ways—operationally vs. creatively—that people do use it. Today’s
post is more about whether we should
use it, and how often, particularly in light of the resources (electricity and
water) that it consumes. Is environmental responsibility a compelling reason to
curb our use of GenAI?
Some housekeeping
As I’ve explained here,
AI is much bigger than the Large Language Model (LLM) chatbots that we consciously
use as the natural successor to Google. We generally speak of AI as a
productivity tool, but a whole lot of AI is devoted to the invisible algorithms
on social media, YouTube, etc. that grab and hold our attention, threatening to
reduce our productivity. I think of this as secondhand AI (like
smoke). Meanwhile, you’re surely hearing a lot of hype about “agentic AI,”
which can supposedly act on its own volition to achieve a goal. At this point
I’m scared of agentic AI and think you should be, too, but that’s another post.
The AI I’m considering here is Generative AI (GenAI), which is the type of chatbot
(e.g., ChatGPT, Gemini, Copilot, Claude) that you feed a prompt to as a way to
research something, or as a way to quickly compose an essay, letter, or
picture. This is how I believe most people think of AI, which is why the terms “AI”
and “GenAI” are so often used interchangeably.
(Note that if you are reading this post long after April of
2026, and there isn’t a single living human not using GenAI, and/or the robots
have taken over and enslaved you, treat this post as a historical artifact. At
least you’ll get a sense of how society initially approached this technology.)
GenAI at work
If you work for a corporation that is clearly embracing
GenAI, providing you a commercial, “walled garden” version of it, and the
training to go with that, adoption is a no-brainer: do as you’re told and embrace
GenAI immediately. My employer is already monitoring my use of it (though they
haven’t said exactly how), showing my compliance on a dashboard. (My “AI Tools Usage” is showing 87% and green.) I could bristle at this, but a) I have always know my use of company assets is monitored, and b) my employer’s expectation that GenAI will make me more
productive is reasonable, as is their expectation that I will be as efficient as possible.
It’s remarkable how quickly all this has changed. I have seen GenAI’s use go from something my colleagues formerly tried (in vain) to hide, to something that my manager will outright ask me
about. When asked, “Did you use AI to help you with this?” I now assume that
the correct answer is a version of “yes.” (This answer is necessarily nuanced.
Both in terms of being honest and articulating my ongoing value as an employee,
I am sure to explain both how it helped and how it fell short of doing the task
for me.) This week my boss tasked me with figuring out how to create a
NotebookLM chatbot specializing in expertly summarized minutes of every meeting anyone on our team attends (or previously attended), which updates its training data automatically. So if our colleague Joe is on vacation we can ask the chatbot, “Why did Joe Blow switch out the vPlan in Blascorp’s EZ-Pluck profile?” and hope to learn the history. I feel like this
assignment would have been unheard of a year ago.
But what if you work for a small business, or have your own?
This is a greyer area, of course. A member of my family is a sole proprietor,
and so far has shied away from GenAI because she’s concerned about becoming too
reliant on the technology. I get her point, and have blogged before (here
and here)
about how doing our own thinking and writing prevents us from falling into intellectual
torpor. But isn’t a tool that legitimately improves efficiency something we ought to rely on? After all, we wouldn’t
even think of trying to run a business without email, a laptop, a smartphone, in
many cases videoconferencing capability, and (depending on the business)
various types of specialized software. All of these tools were new once, and
any small business owner still using a typewriter to generate invoices is
surely a) in the minority, and b) wasting a lot of time. From that perspective,
it’s all but inevitable that any small business owner will ultimately adopt
GenAI for his or her business … so why wait?
GenAI at home
Using GenAI outside the workplace is a more complicated
matter, since it’s not helping put food on the table. I mentioned earlier in
this post that my older daughter has occasionally used it rather frivolously,
such as to punk me. Consider this drawing she had ChatGPT create to memorialize
an accident I had at a hotel pool back in 2024, when I got out of the hot tub
too fast and fainted:
Her prompt for this was, “Can you create an image of a tall
skinny white man feeling faint after leaving a hot tub?” As you can see, the
man portrayed looks more hunky than skinny, and my daughter tried three more
times to get the picture more accurate. Given that these were throwaway efforts
(or would have been had I not used them in an early AI analysis
here),
this was devoting rather a lot of computing resources to a pretty trivial
problem, or shall we say exercise. (Of course part of the point for my daughter
was exploring the early technology; it’s not like she’s stuck with throwaway
art as her primary use case for GenAI.)
On the flip side, her sister won’t use GenAI at all,
somewhat on grounds of intellectual authenticity but mainly due to its
environmental impact. The constant construction of ever-larger data centers is
all over the news, with some shocking statistics thrown around about how much
power and water a single GenAI prompt requires. Today I decided it’s time to
vet this claim a bit, studying the available data and describing it in a
context that could help guide our behavior appropriately.
How much electricity
does GenAI use?
With the help of Claude, because it works better than a
Google search, I did some light research and found some great analysis (here)
on the website of Epoch AI, a nonprofit founded to “help people understand what
is happening in AI from a neutral perspective and grounded in the best possible
evidence.” Epoch AI partners with Stanford’s AI Index, which I’ve come across
in my professional life and seems well respected, as well as the UK’s
Department for Science, Innovation, & Technology, which I trust even more (since
it doesn’t have ties to the tech industry like Stanford does). I must
acknowledge that truly disinterested AI research is hard to come by, because
almost every organization doing serious work in this realm has a business
relationship with it.
So to spread out the risk of misinformation I also put this query to ChatGPT,
which came up with similar numbers but from other presumably trustworthy
sources, including ScienceDirect (which Gemini says
“is considered one of the most reliable and authoritative sources for factual
data in the world”) and Cornell University.
So: Epoch AI, in an article from about a year ago, examined
a widespread previous claim that “an individual ChatGPT query requires around 3
watt-hours of electricity, or 10 times as much as a Google search.” Epoch AI,
leveraging “more up-to-date facts and clearer assumptions,” arrived a the
following conclusion:
We find that typical ChatGPT queries using GPT-4o likely consume roughly 0.3 watt-hours, which is ten times less than the older estimate. This difference comes from more efficient models and hardware compared to early 2023, and an overly pessimistic estimate of token counts in the original estimate. For context, 0.3 watt-hours is less than the amount of electricity that an LED lightbulb or a laptop consumes in a few minutes.
For further perspective: according to
this article, “Google says that its median text
query uses around 0.24 Wh of electricity. That’s a tiny amount: equivalent to
microwaving for one second, or running a fridge for 6 seconds.”
But that’s just text queries. Creating a picture uses
a lot more resources. According to this article by the University of Southern California, using GenAI to create a
picture uses 2.9 Wh—over ten times as much as a text query. I had Gemini come
up with some household use equivalents to give this number some context, and here’s what it came up
with:
- Phone: charges your battery about 19%
- LED bulb: about 19 minutes of light
- Dishwasher: about 14 seconds of a cycle
- Clothes dryer: about 2.6 seconds of a cycle
These seem pretty trivial, but if you consider all the
millions of people using GenAI, it can add up, especially if people get it the
habit of iterating a dozen or so times to get the image just right. (For what it’s worth, I got the cover art for this post in two tries.)
How much water does
GenAI use?
Water is another matter, and very difficult to quantify
because the location of a data center has a lot to do with how efficiently it
can cool all its servers. This“ Washington Post” article documented a study, involving research from the
University of California at Riverside, that found that using ChatGPT to write a
100-word email consumed 519 milliliters of water, which is a little more than a
standard bottle. Obviously that is really high, especially considering how many
people use GenAI and how much that’s growing.
At the same time, as pointed out by this article,
many other industries also use a ton of water, and people don’t seem up in arms
about it: “A single burger takes more than 400 gallons of water to produce; a
humble cotton T-shirt takes more than 700. The United States’ 16,000 golf
courses, meanwhile, each have the potential to use on average between 100,000
to 2 million gallons of water per day. (For comparison, Google says its thirstiest data center in Iowa consumed about
2.7 million gallons per day in 2024; most of the company’s data centers used
substantially less.)”
A less abstract
comparison
To be fair, it’s not like we all sit around eating burgers
all the time; for most of us, that’s a treat. Meanwhile, I would hope most
albertnet readers are enlightened enough to hold out for grass-fed beef, which uses a lot less water to produce. And if you’re like me, you buy a lot of clothing secondhand, which helps mitigate the resources required for your
wardrobe. So what’s a better comparison that can help us frame the
environmental cost of using GenAI? I propose: beer.
(I know what you’re thinking: that’s
my answer to everything.” Well, okay … guilty as charged.)
So here is my thought exercise: how does using GenAI compare
to cracking open a beer? And what is the value of the former vs. the latter?
Obviously this is a wide-open scenario so I’ll narrow it down to how I most
often use GenAI: when researching a blog post.
Here’s what Claude had to say about the electricity required
for a 30-minute research session:
Based on current estimates, a substantive text exchange with an AI like this one — say 20–30
back-and-forth exchanges — is probably in the neighborhood of 5–10 watt-hours
of electricity. Google has reported that after major efficiency gains, the
median Gemini prompt consumed about 0.24 watt-hours, representing a 33×
reduction in energy per prompt compared to a year earlier. At that figure, 30
prompts would use about 7 Wh — roughly equivalent to running a phone for 20
minutes or leaving an LED bulb on for half an hour.
Regarding water use, a Mistral AI lifecycle analysis citied by the Brookings Institution found that a typical 400-token exchange consumes
about 45 milliliters of water—about three tablespoons. Multiply by 30
exchanges and you’re somewhere around 1.5 liters of water—very roughly two or
three bottles’ worth attributable to the 30-minute research session. (This varies
enormously by data center location and cooling method, so we should treat it as an
order-of-magnitude estimate.)
To compare the electricity cost of the GenAI session vs. the
can of beer, I downloaded a spreadsheet-based waste reduction calculator directly from the EPA’s website. It is designed to help consumers like me understand the value of
recycling something vs. tossing it. It calculated that recycling a 12-ounce
aluminum can saves 0.3 kWh—which is roughly 40 times more energy than what’s consumed
by an entire 30-minute GenAI research session. Granted, I often generate a
picture to go with my post, but even if we assume it takes five tries to get it
right, the energy cost of those five
images is still only about one-twentieth of the energy wasted by tossing a
single beer can in the trash. And since this is only the energy cost of recycling, which is less than producing
a can from scratch, these numbers are highly conservative. (Meanwhile, I haven’t even factored in the energy required for brewing and transporting the beer itself.)
Meanwhile, the Water Footprint Network, as described here, estimates a total water footprint of 298 liters per liter of beer—so a standard
12-oz can of domestic beer takes over 100 liters of water to produce. More than 90% of that water comes from the agricultural
supply chain (e.g., growing the barley) while the brewery uses about 6–8 liters
per liter of beer (though a large facility may achieve a 3-to-1 ratio). So my 30-minute
research session uses something like 1–2% of the water embodied in the can of
beer I might have next to my keyboard. (Full disclosure: there’s a now-empty
pint glass on the arm of the sofa as I type this. Yes, drinking while blogging: a
rhetorically risky and planet-impacting combination. So sue me.)
Factoring in value
So that covers the environmental cost of researching a blog
vs. drinking a beer. But what about the value of each? Discounting pub crawls
with my friends—which occur far more
seldom than I would like, to the point that they’re a rounding error—I’m really
talking about unwinding with a solitary beer at the end of the workday. So in
general the value of that beer accrues solely to me.
So does my blog-related GenAI research create any value to
justify its water and electricity use? In the interest of humility I won’t
merely assume this, and will instead dive into the data. Pageview stats across
my blog wouldn’t be very representative, as at least half my posts don’t
require any research at all. So for lack of a better idea, I’ve decided to
analyze the pageview count for each of the albertnet posts that are about AI. After all, those have to be
among the most GenAI-intensive of all, because in writing them I was test
driving the various platforms. Here’s a brief summary of how these posts have
performed:
- Total pageviews across nineteen AI posts: 15,578 (so far)
- Average pageviews per AI post: 819.9
- Average pageviews per AI post per month: 35.5
I could conclude that, from a somewhat abstract viewpoint,
each post is seen by a person a day. But averages aren’t very reliable, and greater
specificity is more revealing. Lurking in that “average pageviews per AI post
per month” is a bit of (GenAI-performed) number crunching, accounting for the
fact that the posts that I published years ago have had a lot more time to
accrue pageviews. Ranking my AI posts by pageviews per month shows that they
are gaining in popularity, with the more recent ones averaging two to three
views per day. Here’s the ranking of all these AI posts over time, so you can
see the momentum:
| № |
Views/Mo |
Total Views |
Title |
| 1 | 102.5 | 1,742 | Tech Check-In – How Good is the Latest A.I.? – Part II |
| 2 | 85.7 | 257 | New Year's Resolutions — AI Edition |
| 3 | 82.8 | 1,077 | What Is ChatGPT Great At (and Not)? |
| 4 | 69.9 | 1,189 | Tech Check-In – How Good is the Latest A.I.? – Part I |
| 5 | 62.4 | 312 | AI Smackdown – ChatGPT vs. Copilot vs. Gemini |
| 6 | 58.0 | 290 | More AI Smackdown – ChatGPT, Copilot, & Gemini Write Poetry |
| 7 | 51.2 | 256 | Tech Reflection – Two Sides of AI |
| 8 | 27.4 | 1,040 | A.I. Smackdown – English Major vs. ChatGPT – Part 2 |
| 9 | 27.1 | 1,031 | A.I. Smackdown – English Major vs. ChatGPT – Part 1 |
| 10 | 23.0 | 597 | Will A.I. Steal Our Jobs? |
| 11 | 20.0 | 739 | Schooling ChatGPT |
| 12 | 11.1 | 719 | Could Artificial Intelligence Replace Writers? – Part 1 |
| 13 | 10.6 | 680 | Could Artificial Intelligence Replace Writers? – Part 3 |
| 14 | 10.0 | 1,230 | A.I. Smackdown – Moto vs. Cortana vs. Siri |
| 15 | 8.8 | 563 | Could Artificial Intelligence Replace Writers? – Part 2 |
| 16 | 7.3 | 1,201 | Almost Intelligent – Part I |
| 17 | 6.3 | 838 | Smartphones & Artificial Stupidity |
| 18 | 6.2 | 1,016 | I, Chatbot |
| 19 | 4.9 | 801 | Almost Intelligent – Part II |
It would be reasonable to conclude that the more recent
posts, which leverage more GenAI research, are reaching more readers, thus
providing a better ROI. Of course I can’t account for all the possible reasons these posts are more popular, but I reckon that to some degree it’s because of the better use of GenAI. Using this tool won’t make be a better writer, but I’ve always been pretty lazy about research and there’s no doubt GenAI helps there. And whether or not this ROI calculation is completely airtight, I hope this helps you at least appreciate my effort to weigh my GenAI “footprint” against its value.
The bigger point here is that the can of beer is consumed
once, quickly,
leaving nothing behind (except maybe a nice belch). In contrast, the energy
that goes into researching a blog post has an effective cost-per-view
that keeps dropping every month it’s up, in perpetuity. If you use GenAI to draft
an email, how many people will it reach, and low long is its tail? Could you
have drafted it on your own—thus exercising your brain—or did you really need
GenAI?
I’m not trying to imply that only bloggers should use GenAI;
this is just one illustration of a cost/benefit analysis of the use of this
tool. If you are doing something useful and an AI chatbot is helping you do it
better or more efficiently, then it’s arguably worth the energy and water—or,
at least, is a more worthy use of it than shopping for a bunch of clothes, going
out for a burger, and then having a few beers.
The point is to be aware of the environmental cost of this
technology, the same way so many of us do when we decide among driving, biking,
walking, or taking mass transit somewhere.
Just because GenAI takes less water than beef or cotton doesn’t mean we should
ignore its environmental cost, since it’s a whole new way people are consuming
energy and water. As recently as three years ago, almost nobody was using GenAI
in their daily lives; now, it’s an increasingly entrenched behavior, data
centers are expanding rapidly, and in some regions power grids are struggling to keep up with demand.
This being said, I truly don’t believe opting out of GenAI
is the solution; just reflecting on how much it helped me write this post, I
can’t imagine not taking advantage of it. Instead, I’d like to see the millions
of people already using it stop acting like it comes without a cost. It’s the
same as driving: did I really need to surround myself with two tons of steel and
burn a cup of gasoline just to travel a mile to the gym and back? (That was a rhetorical question. I always bike to the gym.)
Speaking of cost: one way to keep yourself honest with GenAI
is to not pay for it. If you are on
an unpaid account and use up your tokens, so that your chatbot cuts you off for
some number of hours, maybe that should be your indication that you’ve gone
overboard. Come to think of it, video games, YouTube, and social media should
have that “feature.”
A final note on GenAI
at work
Now that I’ve examined the environmental cost of GenAI, it’s
worth pointing out a final wrinkle: using it in the workplace is actually much
more efficient than using it at home. Corporations get the most benefit out of GenAI
through Retrieval Augmented Generation (RAG), which is where, instead of asking
a large language model to answer from its entire trove of training data, the
GenAI retrieves relevant documents from a corporate knowledge base (contracts,
manuals, research reports, emails, whatever the organization has indexed), then
passes those retrieved chunks to the model as context for its answer. Tools
like NotebookLM, most enterprise Copilot implementations, and corporate
deployments of models like Gemini or Claude typically work this way.
This is much more efficient than “raw” GenAI like consumers
use. The retrieval step is computationally cheap—essentially a sophisticated
search. The generation step is shorter because the model doesn't have to work
as hard to “remember” or construct relevant context; it’s been handed it. And
the answers tend to be more accurate and require fewer iterations, which means
fewer wasted queries. For a user to opt out of using it on environmental grounds makes little sense,
because the big resource expense has already been incurred. As Claude puts it:
The infrastructure cost of a corporate RAG
deployment is largely fixed relative to usage. The vector database has to stay
current whether 500 employees query it or 5,000. The embedding pipeline runs
continuously. The API connections to the underlying model are on retainer. So
each additional active user essentially dilutes the per-capita environmental
and financial cost of that overhead. An employee who declines to use the tool
isn’t reducing the infrastructure footprint; they’re just reducing the output
derived from it. In accounting terms, they’re lowering the return on a sunk
cost.
Synthesis
Wow, I just threw a ton of words at you, didn’t I? Maybe I’m
the most verbose Large Language Model since, well, ChatGPT! Anyway, here’s my
final conclusion: of course you
should use GenAI. It’s an amazingly powerful tool, and it’s getting better all
the time. Now that it’s here, declining to use it makes about as much sense as
blending a smoothie with a knife and a whisk, or doing arithmetic with an
abacus, or churning your own butter. But use GenAI judiciously. Ask yourself:
is this improving the quality or efficiency of my output? Or am I just being
lazy?
Other albertnet posts on A.I., in order of publication
—~—~—~—~—~—~—~—~—
Email me here. For a
complete index of albertnet posts, click here.
