AI for Sustainability ≠ Sustainable AI

We build AI for people who work in sustainability. That makes one question hard to put down: can we use AI to advance sustainability while ignoring the sustainability of AI itself?

Our answer is no. AI for sustainability and sustainable AI are two different things, and the first does not deliver the second.

The distinction is not ours; it was drawn in 2021 ^[1]. AI for sustainability points the technology at goals like the SDGs. The sustainability of AI concerns the footprint and fairness of the technology itself, across its whole lifecycle. The data published since then show how far apart the two can drift.

The benefits are a forecast. The costs are an invoice.

AI's sustainability benefits and environmental costs sit on two different ledgers. The benefits are largely modelled, future-oriented and conditional. The costs are partly measured, while other impacts remain undisclosed or inconsistently reported.

The strongest benefit estimate is a 2025 study ^[2]: AI could cut 3.2 to 5.4 billion tonnes of CO₂e a year by 2035 across three sectors, power, meat and dairy, and light road vehicles. It is a modelled potential, not a measured result. The authors call it a potential that depends on active policy, not an outcome markets will deliver alone.

The cost side needs no scenario.

Both companies have 2030 climate targets, and both name data-centre growth among the pressures on them. As AI infrastructure expands, investors need to see how that growth affects progress toward those targets. Netting a projected global benefit against an actual local cost is the accounting move sustainability professionals reject in the companies they assess. We should not ask them to accept it from us.

Efficiency is real, and it is being outrun

The usual reply is that AI is getting cheaper per query, fast. That is true. Google reports a median Gemini text prompt now uses 0.24 Wh, 33 times less energy than a year earlier ^[6].

In the same year, Google's electricity use rose 37% ^[4]. The IEA sees the same tension across the sector: energy per task is falling quickly, yet total data-centre electricity demand is projected to roughly double, from 485 TWh in 2025 to 950 TWh in 2030 ^[3]. Falling energy per task does not guarantee falling total use.

This is the rebound effect described in 2025 research ^[7]. Cheaper intelligence gets used more, and in heavier forms such as agents, reasoning models and video. Efficiency lowers the cost of each task without limiting how many tasks we run.

Per-query numbers also hide a lot. Google's figure is a self-reported median that has not been independently verified. Mistral's full lifecycle assessment ^[8] counts training and about 18 months of operation, and lands at 1.14 g CO₂e and 45 mL of water per 400-token response. The two measure different things, a median prompt's inference and a lifecycle that includes training, so they should not be compared directly.

Sustainable also means trustworthy

Energy and carbon are only half of it. Sustainability has always included the social and the governance questions, and AI has its own versions of both.

In sustainable finance the sharpest one is evidence. A language model can summarise a company's sustainability report fluently and still misread the evidence. It can also reproduce a company's claims accurately without establishing whether they reflect what the company does. Research published at ACL 2025 ^[9] examined how language models handle exaggerated sustainability claims and points to the risk that they repeat those claims instead of testing the evidence behind them. Extracting a disclosure is not the same as verifying performance.

An AI system that presents unverified sustainability claims as established fact can amplify greenwashing instead of detecting it. Without traceable evidence and appropriate validation, its conclusions are harder to audit and its part in an investment decision is harder to defend.

Regulators have noticed. OSFI's Guideline E-23 on model risk ^[10] takes effect on 1 May 2027 and explicitly covers AI models, including those bought from vendors. It applies to federally regulated financial institutions, in proportion to their risk. For Canadian banks and insurers, the AI inside a sustainability tool can become a model they must govern.

"We are too small to matter"

This is the objection we take most seriously, because it is partly right. A company that builds on other people's models has a small direct footprint and no say in how those models are trained.

But the choices that remain are the large ones. A 2024 study of 88 models ^[11] found general-purpose generative models far more costly than task-specific ones for comparable tasks. On extractive question answering, the most efficient task-specific models emitted about 0.3 g CO₂e per 1,000 inferences against 10 g for multi-purpose ones, roughly 30 times as much. The ratio depends on the models and task compared.

Which model, how large, how often it is called, where it runs and whether its output is grounded in evidence are all decided at the application layer. So is what gets disclosed. AI services can also feed into a client's Scope 3 inventory as a purchased service, depending on the arrangement and the accounting method. Reliable information about their footprint helps clients estimate it and judge the effect of relying on AI more.

Even clean power is not a free pass. In Quebec, Hydro-Québec says the province's data centres use about 200 MW at peak today and that the sector's peak could reach about 1,000 MW by 2035 ^[12]. Low-carbon electricity cuts the operating emissions of that demand, but not the trade-off: the power still competes with electrifying homes, transport and industry.

Five questions every builder should be able to answer

Sustainable AI for sustainability comes down to five questions. We think any team building AI for this field, ours included, should be able to answer them.

1. 1
   Is AI needed for this task, and which model is appropriate? Use the simplest approach that meets the performance requirement. Extraction, classification and tagging rarely need a frontier model.
2. 2
   Can every material output be traced to its evidence? Each claim, score or flag should link to the passage, document and date behind it, and separate what a company reports from what has been independently verified or inferred.
3. 3
   What is the environmental footprint of one assessment? Energy, carbon and water per unit of work, with the method, boundaries and assumptions published alongside the number. Vendor and region are inputs here: providers and grids differ in what they disclose and in carbon intensity.
4. 4
   How do we monitor the system as models and data change? Track accuracy and consistency over time, review material model or workflow changes, and investigate failures. A system that passed its first test is not reliable forever.
5. 5
   Who is accountable when the model is wrong? Named responsibility for reviewing material outputs, correcting errors and deciding when to change or suspend a system, under a governance standard such as ISO/IEC 42001 or the NIST AI Risk Management Framework.

None of this is exotic. It applies to AI the discipline sustainability professionals already apply to everything else: measure, disclose, and show the evidence.

Holding both ideas at once

We believe AI can help capital find better companies and help institutions make better decisions. We would not be building it otherwise.

That belief is a claim about a handprint, the positive impact a technology can enable. A handprint only counts when the footprint beside it is known. AI for sustainability is a statement about what a tool is for. Sustainable AI is evidence about how it was built. Our field should expect both. Read how we handle data and models.