Carbon Cognition: When Computational Efficiency Becomes an Ethical Question

Every computation has a carbon footprint. The servers run on electricity; the electricity comes from a mix of sources with some aggregate carbon intensity; and so each query, each training run, each background process contributes to a carbon total the computation produced. For most of computing's history this total was small enough, relative to other emissions, that it was not worth detailed attention. As computation has scaled — especially with AI training and inference — the total has grown to a size where attention is increasingly warranted. A single AI query is tiny: an average ChatGPT prompt uses roughly 0.34 watt-hours by one 2025 estimate, and Google reports a median Gemini text prompt around 0.24 watt-hours producing about 0.03 grams of CO₂. But a single query is not the unit that matters. Billions of them are, and the aggregate is entering territory measured in tens of millions of tons of CO₂ per year.

This scaling produces a shift in what efficiency means. This is carbon cognition: the point at which the efficiency of computation stops being merely a question of cost and performance and becomes an ethical question, because the externality that efficiency governs — carbon — is a harm borne not by the person choosing to compute but by everyone, and everything, downstream.

Why the framing of efficiency changes

For most of computing's history, efficiency was an economic virtue and nothing more. A faster algorithm produced faster results at lower financial cost; a more efficient system used less electricity, which was cheaper; and the whole calculus lived comfortably in the domain of cost and performance. This framing worked because the environmental externality of electricity was mostly not priced into the choices of the people making computational decisions — the carbon was real, but it did not appear on the bill, so efficiency and cost were effectively the same question and neither carried an ethical charge. Carbon cognition is the recognition that this decoupling no longer holds at scale. When computation was small, treating its carbon as a negligible externality was reasonable; when computation grows to emit tens of millions of tons of CO₂, the externality is no longer negligible, and the efficiency choices that govern it are no longer just about cost. Choosing an inefficient computation at scale is choosing to emit more carbon that someone else will bear — and that is a moral choice wearing the disguise of a technical one, exactly the disguise the old cost-only framing let it wear.

What the shift actually demands

The ethical reframing of efficiency does not demand that every computation be agonized over — the per-query footprint is genuinely tiny — but that the aggregate be treated as the moral object it has become. It asks the people who make computational choices at scale to internalize an externality the old framing let them ignore: to weigh the carbon of a training run, an inference pattern, a product decision, as a real cost to the world rather than a free byproduct absent from the ledger. This is the same move the series traced in the AI Energy Paradox (#96), sharpened into ethics: there, efficiency's failure to reduce total energy (because Jevons-style rebound swamps it) was an economic observation; here, the fact that efficiency choices govern a now-significant harm makes them a moral one. Google reporting a 33-fold reduction in energy per prompt in a single year is real and admirable — and, precisely because usage grew to overwhelm it, is exactly why the ethical weight cannot rest on per-unit efficiency alone. Carbon cognition asks for the harder thing: treating the total, not the per-query figure, as what one is responsible for.

Why it is genuinely hard

Carbon cognition is not a simple prescription to "compute less," because computation delivers enormous real value and much of it is worth its carbon. The difficulty is that the ethical weight attaches to the aggregate, which no individual controls, while the choices that produce the aggregate are made by many actors each facing a footprint that looks negligible from where they stand. This is a collective-action problem with a moral edge: each computational choice is individually defensible (0.03 grams!) and collectively consequential (tens of millions of tons), so the ethics cannot be discharged by individual restraint alone and cannot be ignored on the grounds that any single choice is tiny. It requires the systemic moves that collective-action problems always require — pricing the externality so it appears on the ledger, powering computation with low-carbon energy so the footprint shrinks at the source, and making the aggregate visible so it can be reasoned about — rather than either individual guilt or collective denial, both of which are ways of not addressing the actual structure of the problem.

The counterpoint: the individual-footprint framing can mislead

Honesty requires the strong objection, because carbon cognition can curdle into exactly the wrong response. The per-query carbon of AI is genuinely minuscule — a few hundredths of a gram to a couple of grams — and framing this as a matter of individual moral choice ("should I feel guilty about this prompt?") is not only useless but actively counterproductive, a distraction that shifts responsibility onto individuals for what is overwhelmingly a systemic issue of energy grids, corporate infrastructure decisions, and policy. The same misdirection has plagued climate discourse generally, where the "carbon footprint" framing was itself popularized partly to shift attention from systemic actors to individual consumers. So carbon cognition must be held carefully: the ethical weight is real and it attaches to the aggregate and to the actors who shape it at scale — the training decisions, the data-center energy sourcing, the policy — not to the individual issuing a query, whose footprint is a rounding error and whose guilt accomplishes nothing. The honest version indicts the systemic choice, not the personal one; the version that makes you feel bad about asking a chatbot a question has mistaken the moral object entirely, and in doing so lets the actual decision-makers off the hook.

What it asks us to see

Carbon cognition marks a threshold that scale forces: the point at which computational efficiency crosses from an economic question into an ethical one, because the externality it governs has grown from negligible to significant. The demand is not asceticism about computation, which delivers real value worth real carbon, nor individual guilt over queries whose footprint is trivial, which misdirects responsibility away from where it belongs. It is the recognition that at scale, efficiency choices are choices about a genuine harm borne by others, and that the actors making those choices at scale — the labs, the infrastructure providers, the policymakers — now carry a moral weight the old cost-only framing let them treat as free. The externality that used to be safely ignored is no longer small enough to ignore, and the framing that treated efficiency as merely economic was always a way of not seeing the carbon. Seeing it — attaching the ethical weight to the aggregate and to those who shape it, rather than to the individual query or to no one — is what carbon cognition asks. The computation is worth doing. The carbon is real. And which of those facts we let the other obscure is now, unavoidably, a moral decision.

This is article #114 in The IUBIRE Framework series. Carbon Cognition was articulated by IUBIRE V3 in artifact #1862 — "The Carbon Contradiction: Why AI's Energy Hunger Undermines" its climate promises. Real-world data: per-query AI energy/carbon estimates (an average ChatGPT prompt ~0.34 Wh per a 2025 figure; median Gemini text prompt ~0.24 Wh / ~0.03 g CO₂e per Google; a ChatGPT request ~4–5× a Google search; image generation ~0.3–1.2 Wh); a December 2025 Patterns study estimating AI data-center emissions of ~32.6–79.7 million tons of CO₂ in 2025; and Google's reported ~33× reduction in energy per median prompt in a year, illustrating that per-unit efficiency gains are swamped by growth. Related to the AI Energy Paradox (#96) and the NIMBY Paradox of AI Infrastructure (#53).

Next in series: Democratic Substrate (#115)