The efficiency trap

A seed scenario for the Epistemic Futures Atlas. Counter-imaginaries are warnings — futures we describe in order to avoid them.

In this future, “sustainable AI” won. Every provider reports emissions per inference, and every number improves every year. Yet data-center energy demand triples, water-stressed regions host ever-larger facilities, and rebound effects swallow every efficiency gain. Reporting frameworks proliferate, but they measure what is easy rather than what matters: the direct footprint of compute, never the downstream effects of what the compute does — the consumption it stimulates, the labor it displaces, the attention it captures.

Sustainability has been operationalized as a constraint to be minimized, not a logic to build from. Nobody lied; the metrics were simply too narrow.

Questions this scenario raises for the present:

• Which of today’s AI sustainability metrics would fail to detect this future?
• How can assessment tools capture downstream (Scope 3) impacts without becoming impossibly heavy?
• What institutional teeth keep disclosure from becoming ritual?