The AI–Sustainability Paradox

Model

The framework treats every AI system as a single object with two attributable emissions streams: the infrastructure that runs it (training, inference, cooling, embodied hardware) and the change it produces in the sectors it serves (efficiency gains, avoided emissions, induced demand). It insists that both streams appear on the same ledger before any ‘net’ claim is made. The framework is not a verdict on AI; it is a discipline for evaluating specific deployments, and its usefulness comes from making certain kinds of rhetorical moves — advocating gains without costs, or costs without gains — visible as omissions rather than arguments.

Components

1. Component 1

   Infrastructure Ledger

   All emissions and material flows attributable to running the system: training energy, inference energy, embodied hardware footprint, cooling water, siting-related grid effects.

2. Component 2

   Downstream Ledger

   Measured, not modelled, changes in the sectors the system serves — energy, materials, waste, land use — attributable to its deployment.

3. Component 3

   Rebound Term

   Induced demand and displacement effects: efficiency gains that lower unit cost and expand consumption, or that shift load to less-visible actors.

4. Component 4

   Accounting Boundary

   The explicit statement of whose books close the ledger. Without an owner of the boundary, gains and costs remain on separate documents that never sum.

Use cases

• Regulatory review of ‘AI for climate’ programs seeking to count toward emissions targets.
• Corporate disclosure design for AI-enabled products claiming sustainability benefits.
• Grant portfolio evaluation for climate-tech funds deploying AI-heavy interventions.
• Journalistic evaluation of vendor claims that lack an infrastructure line item.

References

• IEA — Energy and AI (2025)
• UNEP — State of Finance for Nature (2026)