LANDMARK · 7 MIN · JUDGE
Labour and sustainability
After this landmark, you can account for the human labour and the energy behind an AI system honestly — without either ignoring them or using them as a reason to dismiss the tools outright.
A model feels like pure software: you type, it answers, nothing is consumed. Two things are hidden by that feeling. The first is labour. Models are shaped by large amounts of human work — data collection, annotation, and safety rating, including people paid to review disturbing material so that users don’t see it, often through subcontractors and often poorly paid. The second is physical cost. Training a large model consumes substantial electricity, and serving it consumes more in aggregate, because inference happens millions of times a day; data centres also draw on water and grid capacity that is local to somewhere specific. Neither fact settles the question of whether to use AI — a video call has a footprint too, and so does the flight it replaced. What they do is make the honest framing possible: these are real costs borne by identifiable people and places, and they belong in the decision rather than outside it.
This is the canonical concept. It stays the same across learner lenses so personalization never changes the underlying facts.
◌ EVERYDAY LENS
What this looks like for you
Scale your usage to the job. Asking a large reasoning model to do arithmetic or look up a fact you could have searched is real energy for no gain. It’s also worth knowing that the smooth, safe experience you get was partly produced by people who were paid to look at the worst of the training data.
MAKE A DECISION
A team proposes running a large reasoning model over every incoming support ticket to tag its topic — roughly 40,000 tickets a month. What is the soundest response?
CARRY THISTake one repeated AI task in your work. Estimate how many model calls it makes in a month, and write down whether a smaller model, a cache, or no model at all would produce the same outcome.