Orignally posted on amarpandey.me
We are making an expensive, resource-intensive technology artificially accessible and encouraging mass adoption before we've established that its widespread use is economically and environmentally sustainable.
Something about the current wave of AI adoption doesn't sit right with me.
Every product is adding AI features. Existing products are being rebuilt around it, and developers keep finding creative ways to put it into problems that were already solved quite well.
I get the excitement. AI is genuinely powerful. But I think we're getting carried away.
I recently saw an expense-sharing app demo a voice assistant. You say, "Paid 400 for dinner, split into 4," and it creates the expense.
How much better is that than typing into a few fields? An amount, a description and a number of people is a simple data-entry problem. Traditional software handles it perfectly well. The voice feature looks great in a demo, but it adds complexity and compute cost without saving the user much of anything.
You'll find the same pattern in task managers, note-taking apps, calendars and analytics tools. Sometimes I wonder whether we're making products better or just making them look more impressive.
Just because we can build something with AI doesn't mean we should.
Are we solving problems or just finding places to use AI?
Tax calculations, input validation, payments, balances and rate limiting all have predictable inputs and well-defined outputs. Conventional software solves them faster, cheaper, more reliably, and in a way you can actually debug.
Why add an expensive, probabilistic component when a deterministic one already works?
That doesn't mean AI has no place in these products. Scanning a receipt, pulling line items out of a restaurant bill, or answering questions about your spending are genuinely good uses. They make hard things easy.
But using AI to solve a problem that was difficult is very different from using AI to replace an interaction that was already simple. I think we've started confusing the two.
The question shouldn't be how much AI we can fit into a product. It should be whether AI makes that product meaningfully better.
Accessible isn't the same as sustainable
AI isn't free, even when the price makes it feel that way.
Every inference runs on specialised hardware, electricity and infrastructure. Yet AI is being handed out through cheap subscriptions, generous limits and aggressive API pricing, funded by an enormous infrastructure race. The price users pay doesn't necessarily reflect what it costs to produce, let alone its environmental cost.
There's a difference between a technology getting cheaper through better engineering and a technology being made cheap to accelerate adoption.
This matters for anyone building products. Traditional software can serve more users without its costs growing at the same rate. A product that calls a model on every interaction has costs that grow with usage. Put AI into interactions that never needed it and you've added cost without adding value. You can have a great product, growing adoption, and still struggle to make the business work.
Cheaper AI doesn't mean less consumption
AI is getting more efficient. Models are smaller, infrastructure is better, and the cost of a given level of performance keeps falling.
But as it gets cheaper, we find more reasons to use it. Tasks that never justified a model call suddenly do. Features with marginal value get shipped anyway.
Economists call this the Jevons paradox: when something becomes more efficient to use, total consumption can rise enough to cancel out the gains. The International Energy Agency describes exactly this: energy per AI task is falling sharply, while total electricity use from AI data centres keeps climbing.
So are we making AI more efficient, or just making it possible to consume a lot more of it?
The environmental cost we're not talking about enough
AI isn't just software running in the cloud. It runs on data centres, chips, cooling systems, power grids and water, and every one of those has a physical footprint, from manufacturing the hardware to disposing of it when it's obsolete.
A single text query isn't an environmental disaster. But billions of them, plus image, video and agentic workloads that cost far more, add up. And we're building new infrastructure as fast as we can to keep up with demand.
I'm not saying AI isn't worth these resources. I'm saying we should ask whether each new use case is. Does another AI summary, another voice interface, another automated trivial interaction justify what it takes to run?
Sometimes, yes. Quite often, I suspect not.
The middle ground
Most conversations about AI land in one of two camps. The maximalists want an LLM behind every button and text box. The sceptics dismiss it entirely. I don't agree with either.
My view is simpler: use the right tool for the job.
If a problem is deterministic, such as calculations, validation, state or business rules, use deterministic software. If it involves ambiguity, unstructured data, language or reasoning, AI can be an incredible tool, and there's no reason to avoid it.
I don't want to build software without AI. I want to build software that uses AI where it makes sense.
The goose and the golden eggs
I doubt the people steering this, at tech companies or in governments, are unaware of the costs. But countries are racing each other, companies are racing for dominance, and nobody wants to slow down first. Add our fascination with what AI can do next, and it's easy to stop asking what it takes to get there.
It reminds me of the goose that laid golden eggs. Its owner wasn't content with one egg a day, so he killed it to get all the gold at once.
We're so focused on extracting everything AI can offer that we're paying too little attention to what sustains it.
We're not there yet
I believe AI has an extraordinary future in science, medicine, accessibility and much more. I'm excited about it.
But not every app needs to be an AI app, and not every interaction needs an inference call. I'd rather we spend this time making AI genuinely cheaper and more efficient, reducing its energy and water use, and being honest about its real costs, before putting it in everyone's hands by default.
Widespread adoption should follow genuine technological and environmental progress, not precede it.
Perhaps the next big step in AI isn't finding more things to use it for.
Perhaps it's learning where we don't need it at all.
Top comments (0)