How to use AI for good without pretending the cost doesn’t exist.

Paging Erin Brockovitch? 

I want to talk about something that happened in Southaven, Mississippi, because it is exactly the kind of story that makes AI's environmental cost feel real rather than abstract.

xAI, Elon Musk's artificial intelligence company, installed 27 massive methane gas turbines in a majority-Black community in DeSoto County to power Colossus 2, an AI data center just across the state line in Memphis. The turbines went live without federal air permits, which are legally required and exist specifically to force companies to install pollution controls before firing up equipment that spews nitrogen oxides and formaldehyde into the air.

Rev. Robert Tipton Jr., who has battled sarcoidosis for 30 years, describes seeing smog in the air within three miles of the facility. DeSoto County already gets an F grade from the American Lung Association for air quality. Mississippi has the lowest life expectancy and the lowest-ranked healthcare system in the country. And yet.

On April 14, the NAACP Mississippi State Conference and the national NAACP sued xAI with legal representation from Earthjustice and the Southern Environmental Law Center. Since the lawsuit was filed, the number of illegal turbines has grown, now totaling at least 46. On May 6, Earthjustice filed for an emergency injunction to stop operations until proper permits are obtained. At maximum capacity, those turbines could power 400,000 average American homes. About 58,000 people actually live in Southaven. Tipton says xAI chose his community because they saw it as an easy fight. "We didn't have an opportunity to put up a fight." He does now.

This is Environmental Justice 101. When there is no public permit process, there is no public input. The communities that absorb the cost of AI infrastructure are almost never the ones benefitting from it. 

You’re Probably Using the Wrong AI model (and Andrew Winston Can Prove It). 

Andrew Winston is a sustainability strategist and author who has been writing about corporate environmental responsibility for decades. I've been following his Substack closely, and his recent piece on AI's environmental footprint is one of the clearest things I've read on the topic. I'm going to credit him liberally here because he put the framework together better than I could.

His core argument: the big AI platforms, Claude, ChatGPT, Gemini, each have three power tiers that operate like an EV's driving modes. The lightest tiers (Haiku, Flash, Instant) use dramatically less energy than the heaviest (Opus, Pro). Not 15% less, like switching from Normal to Sport in a Tesla. More like 5 to 10 times less processing per task. And most of us default to the heaviest tier for questions that don't remotely require it.

The mental model Winston borrows from Claude itself is useful: don't think about it as simple versus complex tasks. Ask yourself how much does a reasoning error cost me? If the answer is "not much," you don't need the nuclear option.

The AI Tier Guide, in Plain English

Lightest tier (Haiku, Flash, Instant): Use for quick factual lookups, simple definitions, one-step questions. As Winston puts it, using a heavier tier here is "like taking a cab to your mailbox."

Middle tier (Sonnet, Thinking): Use for tasks that require context, multi-step reasoning, or pattern recognition. This handles the bulk of everyday professional work and is the default on most platforms for a reason.

Heaviest tier (Opus, Pro): Reserve this for synthesis across many sources, long documents, or situations where an early reasoning error has real downstream consequences.

Winston is careful to note that the goal isn't always "lightest." If a lighter tier gets it wrong and you re-run it three times, you may have used more energy than one clean Sonnet run. The goal is appropriate matching. The tier you choose can radically increase or decrease the resource footprint of your AI use. Start lower, experiment, and reserve the heavy hitters for when they’re actually earned. Winston’s full piece is worth reading here. 

I Built Something!

I’m getting kind of addicted to Claude Code and all of the brilliant things it can do for me, but was personally concerned about the environmental impact to which I was contributing. The following is a tiny example of how I tried to combat that.

I’ve been thinking about what annoying sustainability decisions might be better automated to make us all live more sustainably. The first one that came to mind was automating carbon offsets after I book a flight, which I forget to do at least 30% of the time. ‍

Here's How It Works

Using Claude and Google Apps Script (which lives inside your Google account, no downloads, no new subscriptions), we built an automation that:

•       Scans your Gmail inbox every Monday for flight confirmation emails

•       Uses Claude to read each email, extract the route, cabin class, and passenger count, and identify the confirmation code

•       Deduplicates, because airlines send about five emails per booking and you don't need to offset the same flight six times

•       Calculates your CO2e using BEIS/ICAO methodology with radiative forcing included

•       Sends you a summary email with the total footprint, a real-world comparison, and a direct link to buy verified Gold Standard carbon offsets

The Gold Standard clean cooking projects we link to replace open-fire cooking with clean stoves in East Africa. Beyond cutting emissions, they improve respiratory health and disproportionately benefit women and children. The co-benefits are real and independently verified.

One real output from my own inbox: three flights over the past week, 1.10 tonnes CO2e total, estimated offset cost around $16. It found a Southwest confirmation buried in a "Your Cleveland Trip is almost here!" subject line. It found the Delta confirmation from two weeks ago that I had completely forgotten. It did all the math I was never going to do. 

A Note on Why Claude Does the Parsing

Airlines don't have a standard format for confirmation emails. United writes "booking confirmation" in the subject. Southwest writes "Your Cleveland Trip is almost here!" Delta does something else entirely. No amount of search filters can reliably catch all of them. Claude reads the body of each email and determines whether it's actually a flight booking, regardless of format. That's the thing that makes this tool work rather than being a fancy filter.

Concerned about your own AI usage? Read about the real cost of your daily AI habits.