How I Use AI, and What It Costs the Planet
I use AI in my work. I want to be straight with you about how, and about what it costs.
Some people hear “AI” and picture a data center the size of a town, drinking a river. Others hear it and picture a chatbot making things up. Both reactions are reasonable, and neither one is quite what is happening when I sit down to work on your systems.
So this page does two things. It tells you exactly what I use and what I do not, and it puts real, sourced numbers on the environmental side instead of asking you to take my word for it.
The short version
9 sec
of television is the energy in one prompt
11 kg
CO2e in a very heavy week, training included
That is roughly 40 miles of driving. Not nothing, and not a river.
What I actually use
My work runs on Hermes, a personal AI assistant built by Nous Research, set up for me by Tyler through his small business Ampersand. Tyler is a local builder who helps people like me get an assistant configured and working instead of spending six months reading documentation.
What it does
Drafts and edits my writing. Builds and updates pages on this website. Organizes spreadsheets, forms, and email campaigns. Wrangles the logistics of events with hundreds of moving pieces. Checks my work and catches the mistakes I would have shipped.
What it does not do
It does not decide what I offer, how I price it, or what I tell you. It does not publish anything to this site or send any email without me reviewing it. It does not book or cancel your appointments on its own.
Your information
Client details stay in the tools I already use, like my booking system, my email, and my spreadsheets. Intake answers are read by me. I do not feed client information into public AI tools, and I never ask for passwords in a form.
The environmental question, honestly
AI does use electricity. Pretending otherwise would be dishonest. The fair question is how much, and what to compare it to, so here is the real number from the people who actually run the hardware.
Google publishes its own measurement for a median text prompt: 0.24 watt-hours of energy, 0.03 grams of CO2 equivalent, and about five drops of water. That is roughly nine seconds of watching television. It is a measured disclosure with a published methodology, not a guess.
Other researchers get a larger number by spreading the cost of training the model across every query, which comes out around 2.2 grams of CO2 per query. That is the conservative end. Both numbers are legitimate answers to slightly different questions, so everything below shows both.
0.03 g
CO2e per prompt
Google’s own median disclosure
2.2 g
CO2e per prompt
training costs included
0.24 Wh
energy per prompt
about 9 seconds of television
What a week of my AI use looks like
A heavy week for me is somewhere between 250 and 5,000 prompts, counting every draft, revision, page edit, spreadsheet pass, and tool call. Here is what that costs, using both published figures:
| In a week | Google’s figure | Training included | Rough equivalent |
|---|---|---|---|
| 250 prompts | 8 g | 550 g | 0.0 to 3 km of driving |
| 1,000 prompts | 30 g | 2.2 kg | 0.2 to 13 km of driving |
| 2,500 prompts | 75 g | 5.5 kg | 0.4 to 32 km of driving |
| 5,000 prompts | 150 g | 11.0 kg | 0.9 to 65 km of driving |
Read that bottom row again. Even on the high end, with training costs included, a very heavy week of my AI use comes out near 11 kilograms of CO2. That is about the same as driving 65 kilometers, or a bit over 40 miles. Not nothing. Also not a river.
What it is similar to
Numbers in grams are hard to feel. These are the comparisons that made it click for me:
One prompt
Less energy than a microwave running for one second. About the same as watching nine seconds of television.
1,000 prompts
Between 30 grams and 2.2 kilograms of CO2. Somewhere between a third of a kilometer and thirteen kilometers of driving.
Water
Google measures about five drops per prompt. A thousand prompts is about a quarter of a liter.
For scale
The average American drives roughly 50 kilometers a day. My AI use in a heavy week is comparable to that, not to a data center’s annual draw.
What this does not cover
Being fair means naming the limits. These figures cover inference, which is the part that runs when you actually use it. Training a large model is a genuinely large one-time investment, and that cost is why the second column exists. This page also does not count the emissions of the devices we all use, or the wider internet, because those exist whether or not I draft a page with help.
What I can tell you is that I use these tools for operational work in short, purposeful sessions, not for generating volume. I am one person with a laptop, using a small amount of a shared system to do work that would otherwise take me three times as many hours, and often three times as much paper, printing, and driving.
What I do about it
Using less is only half of a real answer. Here is where I put money and effort to balance the rest.
$100 every month to Junglekeepers
A registered 501(c)(3), EIN 88-1863507, protecting rainforest in the Madre de Dios region of Peru. More than 160,000 acres secured against roads, logging, and extraction.
Through the end of 2026 every dollar is matched by a major donor, toward securing tens of thousands of additional acres that would block an advancing road. My monthly gift is being doubled along with everyone else’s.
Two decades of service
Hundreds of hours given over the last twenty years to organizations that conserve water, clean up rivers, and teach river safety. Fundraising for local nonprofits serving people without homes, veterans, hunger relief, community gardens, and animal support.
Community education on safe and alternative use of plant medicines, and building a local wellness directory so healthy practices are easier to find. This is the work the AI is helping me do more of, not less of.
Solar panels and LED lighting
Our roof generates its own electricity, and every fixture in the house is LED. An LED uses roughly 75 to 80 percent less power than an incandescent bulb for the same light.
One bulb is a rounding error. Twenty bulbs, every evening, for years, is not. The panels cover the daytime draw and the LED fixtures shrink the load that has to be covered, so the house asks less of the system from both ends at once. Small decisions, made once, that keep paying out long after the receipt is gone.
Working remotely, by design
My work happens from a laptop, not a car. No commute, no driving to meetings. The average American drives about 13,000 miles a year, most of it getting to and from work.
For clients the saving is larger. A nonprofit that would have shipped binders or flown someone in gets the same work over a link.
Digital first, paper last
Forms, schedules, dashboards, and handbooks all live online. I have run a conference for two hundred people with almost no paper, name badges being the only thing that had to exist physically.
Paper is never just the sheet. It is toner, printer, shipping, and the reprint when a name is spelled wrong.
Small and purposeful use
I use these tools for specific tasks in short, deliberate sessions. No leaving them running, no generating for the sake of volume, no asking the same question five times.
Restraint is the cheapest offset there is. Every prompt I do not send costs nothing, and no donation has to make up for it.
Why these count. I am not going to hand you a number and claim my donation erases exactly that many grams. Rainforest carbon accounting does not work that cleanly, and anyone promising precision is selling something.
But that does not make any of this symbolic. A monthly gift to permanent land protection. A house that makes its own power. Lights that sip instead of gulp. A business with no commute and almost no paper. Twenty years of given hours to water, rivers, and the people around me.
Those are not offsets on a spreadsheet. They are the thing itself: less taken, and more put back. My AI use is small, my footprint is kept deliberately low, and the money and the hours go out every year whether or not a laptop was open that week.
Where the numbers come from
Google, Measuring the environmental impact of AI inference (2025), the official per-prompt disclosure and methodology, also published as arXiv:2508.15734. cloud.google.com
Nature Scientific Reports (2024), per-query estimates that amortize model training. nature.com
How Hungry is AI? Benchmarking Energy, Water, and Carbon of LLM Inference (2025). arxiv.org
Hannah Ritchie, What’s the carbon footprint of using ChatGPT or Gemini? (August 2025), a clear walk through the competing estimates. substack.com
Car comparison uses roughly 170 grams of CO2e per kilometer for an average petrol car. Driving distance assumes an average of about 50 kilometers per day in the United States.
Questions about any of this?
If you want to know how AI shows up in a specific piece of work, ask me directly. I would rather answer than have you wonder.
And if you want this working inside your own business, that is part of what I build. Getting Started explains the first step.