British grid operator halves solar forecasting error with AI

If you sell AI as a way to cut emissions, model performance alone is not enough. You need to show what changed in the customer's operations: how much fuel they stopped burning, how much electricity they saved, and what the computing cost. Solar generation forecasting offers a straightforward example. The more accurately a grid operator knows how much power solar panels will supply, the better it can plan the operation of other sources.

In an MIT Technology Review opinion piece, Mary Nichols and Jeremy Tamanini examine how far promises of AI's climate benefits are supported by actual deployments. Their conclusion: some solutions work, but there is still too little public evidence to consider the sweeping projections substantiated. For founders and product teams, there is a useful lesson here: between a good model and a measurable impact lies work that a presentation cannot do for you.

Where AI is already helping

The clearest example in the article is the collaboration between Britain's National Grid ESO and Open Climate Fix. According to the data cited by the authors, the project halved solar generation forecasting error. They link this to avoiding hundreds of thousands of tonnes of carbon dioxide emissions a year.

The mechanism is practical. The operator needs to balance electricity production and consumption in advance. Cloud cover changes, solar plants produce varying amounts of power, and forecasting errors must be offset by other sources. A more accurate forecast helps make use of the available solar generation. However, halving forecasting error does not in itself mean halving emissions: these are different metrics, and the relationship between them must be established.

There are examples in transport, too. According to an independent study described in the opinion piece, Didi Chuxing's route optimization reduced emissions per trip by an average of 16%. The causal chain is clear: fewer detours mean less distance traveled and lower fuel consumption. But a per-trip figure does not yet answer how the service's total emissions changed as the number of bookings grew.

Why a projection for 2035 guarantees nothing yet

The authors discuss a study that estimated AI's potential to reduce global emissions by 3.2–5.4 billion tonnes of CO₂ equivalent annually by 2035. The calculation accounted for AI's own energy footprint. Most of the expected improvements are in electricity, the meat and dairy industries, and transport.

The key word here is “potential.” In the food industry, for example, the calculation assumes that AI will help make alternative proteins tasty and cheap enough for consumers to substitute them for animal products. Finding the right formulation is not enough: production, affordable pricing, distribution, and sustained demand will also be needed.

The opinion piece's own examination of deployments has limitations, too. The second author used Claude Code to analyze public information, after which the results were checked against documents and publications. This is a review of available information, not a comprehensive audit of entire industries. The absence of a public case study does not prove the absence of a project. And finding a successful case does not show how easily it can be replicated for another customer.

Finally, AI can help increase oil and gas production. Counting only beneficial applications is therefore insufficient. Nor can projected emissions from all data centers automatically be attributed to AI alone: these facilities also handle other computing workloads.

What to measure in your own product

The practical takeaway for a team is to agree with the customer in advance on the physical outcome being tested. For routing, this might mean kilometers traveled and fuel consumption under comparable conditions. For a power system, it might mean changes in the use of generation and the associated emissions. Model accuracy remains an intermediate metric.

Next, account for the costs of training and running the model, define the comparison period, and check whether the impact persists after the pilot. Assumptions also need to be documented separately: how emissions are calculated, what data is available, and what changed in the business during the experiment. This approach gives the customer a basis for a purchasing decision and the team an honest understanding of the product's value. According to the opinion piece's authors, these links between deployment and results are precisely what the industry's sweeping climate promises still lack.