The Price of Intelligence · the policy tool

GreenBench

A permission standard for large-scale AI

Large-scale AI should be approved only when its ecological burden is auditable, it is likely to outperform realistic alternatives, and it is compatible with local ecological justice. Run a deployment through the three gates below.

Why a standard

Disclosure is not justification

For years, efficiency held data centre electricity roughly flat. AI reversed that: the IEA put data centres at ~415 TWh in 2024 (~1.5% of global electricity) and projects ~945 TWh by 2030, with AI driving the growth. Today's governance answers with per-prompt figures and annual net-zero certificates — numbers that can't say whether any given deployment is worth it. Thin governance risks blocking high-value AI whose absolute use looks large while licensing low-value systems whose per-interaction numbers look small.

Data centre electricity, 2024
~415 TWh
≈1.5% of global electricity (IEA)
Projected by 2030
~945 TWh
AI is the main driver (IEA)
The unit that matters
The task
nobody buys prompts; they buy completed jobs at agreed quality
The assessment

Three gates

Answer for a specific deployment — here, now, for this purpose. The verdict updates as you go: approve, approve with conditions, or refuse. Approvals justified on sustainability grounds are provisional until a post-deployment rebound audit.

Gate 1 · Integrity

Is the ecological burden auditable?

No deployment should be justified on sustainability grounds unless its claims are reported in a standardised, auditable form. Tick what the operator actually discloses.

Gate 2 · Counterfactual justification

Does it beat the realistic alternatives?

The unit of analysis is a completed task at an agreed quality — including retries, oversight, and downstream correction. Benchmark against three baselines: human work, conventional software, and no action.

What does the deployment mostly do?
Could conventional software or a template do it at acceptable quality?
Representative task (from the benchmark ledger)
Model
Retries per accepted output: 1.5×
Human oversight per task (minutes): 15
—
Gate 3 · Justice under local ecological constraint

Can this place carry it?

An accurately measured, comparatively efficient deployment can still be unacceptable if it worsens scarcity or imposes burdens on communities without commensurate public value and meaningful participation. Annual averages obscure this.

Is the site in a water-stressed basin?
Does it strain the local grid at peak periods?
Public value of the use case
Meaningful participation for affected communities?
⚖️Answer the gates above

    Recommendations

    What adopting this looks like

    1. Procurement: public-sector buyers require a GreenBench assessment before approving or renewing large-scale AI contracts.
    2. Disclosure: governments mandate auditable, site-level reporting for large AI and data centre operators — electricity use, peak-period demand, water withdrawal and consumption, local water-stress context, and both emissions bases, on standardised boundaries.
    3. Presumption: planning and procurement bodies adopt a rebuttable presumption against high-volume generative deployments where the no-action baseline is credible and local ecological constraints are material — rebutted by demonstrated public value, rebound auditing, and mitigation conditions.

    Where this sits

    GreenBench is not anti-innovation: it is a targeted permission standard for large-scale deployments that permits high-value substitution while raising scrutiny for induced demand. Efficiency-led Green AI and carbon reporting remain valuable — but neither tells a policymaker whether a deployment should happen. It is deliberately narrower than full lifecycle assessment: stationary compute is the immediate policy bottleneck, upstream of robotics, and the point where approval decisions bite. The benchmark ledger, the live grid page, and the water lens are the measurement layer this standard runs on.

    Grounding: Masanet et al. 2020 (efficiency plateau) · IEA 2025, Energy and AI · Ostrom 1990 (commons governance) · Hardin 1968 · Locke 1988 & Kant 1996 (legitimate appropriation of shared resources) · D'Ignazio & Klein 2020, Data Feminism (measurement makes burdens visible) · Samuel, Lucivero & Somavilla 2022 · Halsband 2022 · Robbins & van Wynsberghe 2022.