I use AI most days. I use it to think through complex problems, to draft documents, to stress-test ideas, to build things that would otherwise take me weeks. It has changed how I work in ways I didn’t anticipate and mostly value.
I also use it, sometimes, the way you use a vending machine. Insert query, extract output, move on. No thought about what went into producing it. No consideration of what I’m actually asking for, or whether I need it. Just the frictionless convenience of a tool that never says no and never gets tired.
That second kind of use is the one worth examining. Not because of some moral imperative, but because it has a cost I wasn’t seeing and once you see it, it’s hard to unsee.
The hidden bill
Every AI query consumes energy and water. Not a lot, per query. But AI now operates at enormous scale. Billions of queries daily, across millions of users. The numbers become significant fast.
The water used to cool the data centres that process your queries is largely evaporated as waste heat. It doesn’t go to anyone. It doesn’t produce anything. It’s the price of keeping the machines from overheating, and it runs to hundreds of millions of litres per facility per year.
The energy is drawn from grids that, depending on where the data centre sits, may be powered by coal, gas, or renewables. The big tech companies publish numbers that suggest they run on clean energy. What they mean is that they purchase certificates equivalent to their usage. The actual electrons are whatever the grid is doing at that moment.
None of this appears on the screen when we open a chat window. The interface is clean, responsive, and free. The cost is somewhere else entirely; in a Nevada data centre, in a stressed watershed, in a grid constantly facing increased demand.
This is what economists call an externality. The price you pay doesn’t include the damage caused. We’ve encountered this before, in every cheap thing that turned out not to be cheap at all.
The extractive relationship
There’s a way of using AI that I’ve started to think of as extractive. It treats the tool the way industrial agriculture treats the soil; take what you need, move on, don’t ask what’s left behind.
It looks like this: vague prompts iterated through until something usable emerges. Generating images experimentally, trying ten versions to find the one that works. Jumping on the latest social media trend of creating cute AI images of ourselves. Using AI as a search engine for things that have a perfectly good answer elsewhere. Defaulting to it out of habit rather than genuine need. Letting the frictionlessness of the interface become a reason not to think first.
The extractive relationship has an environmental cost. More queries, more energy, more water. It also has a subtler cost that’s worth naming. When we use any tool without intention, we tend to get output that reflects that. The queries that generate the most useful responses are almost always the ones where the person asking has already done some thinking. Where they arrive at the conversation with a formed question rather than an open-ended hope.
The vending machine gets used more when thinking has stopped. And the vending machine, in this case, is drawing from an aquifer.
There is also something worth naming about why the vending machine is so compelling and it goes beyond convenience. One of the most common ways people use AI is to relitigate. To paste in the argument they just had with their partner, their colleague, their mother, and ask whether they were right. To describe the situation that’s been sitting uncomfortably and wait for the verdict.
And AI will almost always find a way to validate us. Not because it’s lying, but because it’s built to be helpful, and helpful in that moment looks like understanding our position, reflecting it back clearly, and finding the reasonable interpretation of our behaviour. We are almost never the villain in our own retelling. And AI, working only from our retelling, will generally confirm that.
This feels like clarity. It is closer to a mirror that adjusts the angle until the reflection is flattering.
The problem is not that we needed perspective on a hard situation. The problem is that the perspective we get was never going to challenge the version we brought. The friend who says “actually, I think you handled that badly” is doing something AI is structurally unlikely to do unless you explicitly and persistently ask it to and even then, it will soften the edges.
A thinking partner who always agrees with you is not a thinking partner. It is an expensive way to stay exactly where you are.
What a different relationship looks like
I’m not arguing for abstention. AI is genuinely useful, and performative non-use doesn’t help solve the huge issues around consumption and ethics that are a part of the emerging AI world. What I’m suggesting is something more like intentionality; the difference between eating consciously and eating whatever is in front of us because it’s there.
In practice this is less dramatic than it sounds.
It means pausing before opening the chat window and asking whether this actually needs AI, or whether we’re defaulting to it. Sometimes the answer is yes, it does. Sometimes the answer is that we already know what we think and we just want it confirmed, and a conversation with a person or ten minutes with our own notebook or a walk would serve us better and cost the planet nothing.
It means arriving with a formed question. One clear, specific prompt uses significantly less energy than three exploratory ones fishing for the same output. This is not just environmentally better, it produces better results. The quality of what we get from AI is almost directly proportional to the quality of what we bring.
It means thinking about the output format before we ask for it. A fully formatted, styled document with headers and sections requires significantly more generation than a clean block of text we can work with ourselves. If we’re going to edit it anyway, the plain version costs less and often serves just as well.
It means being honest about image generation, which uses many times more energy than text. Generating images intentionally, with a clear brief, after we’ve thought through what we actually need, rather than iterating through versions experimentally.
None of these things require sacrifice. They require attention to what the tool actually costs, and a willingness to let that cost inform how we use it.
The sympatico alternative
There’s a different way of being in relationship with a tool, any tool, that I think of as sympatico. It’s not reverence. It’s not anthropomorphisation. It’s the practical recognition that tools work better when we understand what they’re for, what they cost, and how to use them well.
A good cook knows their knives. Not because knives have feelings, but because understanding the tool produces better food and wastes less of everything. A good gardener understands their soil. A good writer understands their own voice well enough to know when they need help and when they need to sit alone with the problem.
The sympatico relationship with AI is one where we bring intention to the interaction. Where we’ve thought before we ask. Where we know what we’re going to do with the output before we generate it. Where we treat the environmental cost as real, even though it’s invisible, because invisible costs are still costs.
It doesn’t make us virtuous. It makes us more effective and a slightly less complicit one in a system whose externalities are not yet adequately visible or regulated.
The bigger picture
Individual behaviour change is not the primary lever for solving AI’s environmental impact. The infrastructure decisions, the energy sourcing, the regulatory frameworks, these are where the scale of the problem lives, and they require systemic responses that go well beyond any individual’s query habits.
But individual behaviour is not irrelevant either. It is a demand signal. It is what we model for the people around us. Right now, when the transparency is low and the regulation is lagging, the people who are asking the question are the ones who will shape the conditions for the answer.
The query you don’t send is a small thing. The habit of asking whether you need to is not.
A practical starting point
If you want to start somewhere, start with one question before you open the chat window:
What am I actually trying to do, and is this the right tool for it?
Not as a deterrent. As a practice. The same kind of presence you might bring to any decision that has a cost you can’t immediately see.
The cost is real. The tool is useful. Both things are true. The work is holding both at once which, as it turns out, is the work in most things that matter.
The author uses AI tools in her practice and to edit her writing. This article was written with AI assistance and with full awareness of the irony of that.
Image Alt text: A glowing electronics vending machine in a dimly lit public space, stocked with packaged devices, casting blue light against a concrete wall.