Most of What I Changed Didn't Matter

Amy Westlake

4 days ago

5 min read

Situation

Someone posted a question in one of our company Slack channels asking what we were doing about the environmental impact of all the AI use.

I'd been skimming a few articles about data center energy and water for a couple of weeks. It was the kind of reading where you file something under "concerning" and go back to work. But the question sitting there in Slack wasn't about the company, or it was, but that's not how I took it.

What I heard was: what are you doing about it.

And I use AI more than almost anyone I work with. I'd spent months telling other people to use it more. And honestly, I wasn't doing anything.

The AI Move

I opened Perplexity and typed this (I know... the irony of the moment):

I use AI a lot. What can I do to reduce/offset the impact the AI data centers have on the environment? I feel like I'm contributing to the problem.

What came back was generic. Some of it was fine. One suggestion was to join a carbon offset program, and that one annoyed me because I wasn't looking for a way to pay my way out of it. I was looking for something to do differently.

That's on me, though. Vague prompt, vague answer. I should have been more specific with my ask.

It took several rounds of follow-ups before I got to anything I could actually act on. Eventually I had a short list. Shorter prompts. Stop saying please and thank you. Turn off video autoplay. Delete files I'm not using. Uninstall apps I never open.

I changed a few things that week and stopped thinking about it.

The Shift

That was October last year. Eleven months ago.

I went back this week and checked what actually stuck.

Shorter prompts: dead. Completely dead. I use the microphone for almost everything now, and when I talk I don't filter the way I do when I type. Way more words go in. The AI has to ask me fewer questions to figure out what I want, so the whole exchange ends up shorter even though my half of it got longer. I didn't know which of those was better for anything.

Turns out it's the second one, and not by a little. Google published measurements showing that most of the energy behind a single prompt isn't the problem. Idle machines held in reserve, cooling, hardware around the hardware are all overhead. Every time you hit Enter, regardless of how short you keep it, you're paying the cost. Five short messages cost more than one rambling one.

So the habit I didn't mean to drop was the wrong habit to have picked up in the first place.

I still delete files. Relentlessly, actually, and I have no intention of stopping. That one does nothing. Deleting a thousand emails saves roughly five grams of CO2, and the half hour you spend on a laptop deleting them costs somewhere between five and twenty-eight. Storage draws its baseline whether my files are sitting there or not.

The politeness came back within about a week. I say please and thank you to AI because I say please and thank you to people, and I'm not going to run a separate personality for a chat window. At the scale of hundreds of millions of users that's apparently real money in electricity. At the scale of me, it's a few words a day. I'm keeping it.

Then I looked at what I've actually changed in eleven months without ever calling it that.

I run about ten scheduled tasks now. Last week I rewrote three of them, stripped out work that two of them were duplicating, and tightened up how they run. I do that constantly. Not on a schedule, just whenever I notice one is producing something I've stopped reading. The number of tasks isn't the thing. What matters is whether each one is doing exactly what I need at the moment I need it.

I also start a new chat every time the topic changes. Religiously. I picked that habit up because long threads get muddled and give me worse answers, and every turn in a thread drags the entire history back through the model again.

Both of those cut real energy. I made neither of them for that reason.

The Pattern

The changes I made last October were the ones I could make in an afternoon. Delete something, trim something, flip a setting off. Small enough to do immediately, which is exactly why none of them were the ones that counted.

The things that genuinely reduce what I consume are decisions I was already making because sloppy output annoys me. Rewriting a scheduled task that's gone noisy. Killing a chat that's drifted. Capping a search so it stops digging. Every one of those is a quality decision that happens to also be an energy decision.

So the question I asked Perplexity was slightly off. I asked what I could add, like reducing my footprint was a separate activity I'd bolt onto the way I already work. The version worth asking was closer to: where am I being wasteful? Wasteful and expensive turn out to be the same thing most of the time.

There's one I'm still working on. I default to the most capable model and the deepest thinking mode for everything, including questions that need neither. That gap is enormous. Reasoning mode uses somewhere around thirty times the energy of a standard response, because the model generates hundreds of times more text you never see. On tasks that involve actual reasoning, that's worth it. On "summarize this" and "reword that," I'm paying thirty times the cost for nothing. I need to just remember to switch the model.

The Implication

Go find something you changed months ago for a good reason and check whether it's still true.

That's it. Not a new habit. An audit of the ones you already have, because I'd have told you with total confidence that I write shorter prompts now, and I'd have been wrong for most of a year.

If you want somewhere specific to start: look at whether your AI tool has a thinking or reasoning toggle, and notice how often you leave it on for something that isn't hard.

What I'm Testing Next

Nothing, really. Not on this one.

The thing I keep coming back to and not resolving is water. I'd absorbed "AI uses enormous amounts of water" as a flat fact and never localized it. Nationally, data center cooling was under one percent of US water use in 2023. The real problem is where they get built. Arizona, New Mexico, and Southern California already deal with water shortages before adding AI data centers. Texas is projected to go from under one percent to somewhere between three and nine percent of state water by 2040. That's a siting question and a zoning question, and none of my prompting habits touch it. I'll also say the uncomfortable part out loud: knowing all this, I still don't want one built in my town.