• The opinion piece's central finding is that AI is not yet being used at scale to cut emissions, whatever the optimistic projections say.
• Of the grid operators the authors reviewed in China, the US and Europe, only National Grid ESO had a working AI forecasting program.
• The deployment that shipped, National Grid ESO with Open Climate Fix, halved its solar-forecast error, according to the article.
• A Nature Sustainability study estimates AI could cut 3.2–5.4 billion tons of CO2-equivalent a year by 2035, net of AI's own energy use. The authors say real-world deployment is far behind that.
• The IEA projects data-centre electricity consumption of about 945 TWh by 2030, so the footprint is growing faster than the verified savings.

The climate case for AI has not been shipped. In an MIT Technology Review opinion piece published on November 9, 2026, Mary Nichols and Jeremy Tamanini argue that industry is not yet using AI to cut emissions in a meaningful way. They wrote it as countries met in Antalya, Turkey for the UN climate conference, where the host asked for commitments that AI be designed, powered and deployed in support of climate action. The gap between the forecast and the deployment is the story.

The benchmark is a model, not a measurement

The authors anchor their critique on a peer-reviewed Nature Sustainability study from the LSE Grantham Research Institute and Systemiq. It estimates AI could cut 3.2 to 5.4 billion tons of CO2-equivalent a year by 2035, net of AI's own energy use, with gains concentrated in power, meat and dairy, and vehicles. That figure is a scenario estimate, not an observed result, and the article treats it as the ceiling to check deployments against.

Power is the clearest test, and one operator passed

Grid forecasting is where the case is strongest and the deployment thinnest. The authors say AI forecasting could raise renewable delivery from existing solar and wind by up to 20 percent, avoiding an estimated 1.8 billion tons a year. Of the operators they reviewed in China, the US and Europe, only National Grid ESO had a working program. Its work with Open Climate Fix halved solar-forecast error, which the article credits with avoiding hundreds of thousands of tons of CO2 a year.

The review method matters for how much weight to give this. The authors describe an AI-assisted review of public information that they then checked themselves. That is a reasonable way to survey operators, but it is not an audit, and a program that exists but is not publicly documented would be missed. Treat the one-in-the-sample count as the authors' reported finding, not an independent census.

Transport results rest on mixed evidence quality

The transport examples differ sharply in how well they are supported. The article reports that CATL built a battery-design system trained on more than 100,000 design cases and 600 TB of test data, reportedly cutting material-screening time by two orders of magnitude, but that is a company claim. By contrast, independent researchers found Didi Chuxing's AI routing cut per-trip emissions by 16 percent on average. The second result is the one a reader can argue with.

The footprint side is measured, and it is growing

On the cost side, the IEA's Energy and AI report puts data-centre electricity consumption at about 945 TWh by 2030, more than double the current level, with a base case near 1,200 TWh by 2035 and a scenario range of 700 to 1,700 TWh. The article adds that the IEA base forecast puts global data-centre emissions above 300 million tons a year by 2035. That load is projected, but it is being built now, while the savings depend on adoption.

Non-technical bottlenecks and fossil use limit the upside

The authors point to constraints that model quality does not fix. AI does little to shorten certification, permitting, building and marketing timelines, which gate how fast any new product reaches the market. They also cite reporting in Nature that oil and gas firms are pursuing AI, including to boost fossil fuel production. The same capability that finds efficiency can raise output of the thing being decarbonized.

Verdict: judge AI climate claims by deployments, not forecasts

Our reading is that the piece is a fair call to hold climate claims to production standards. A 3.2-to-5.4-billion-ton estimate is not evidence of a shipped system, and one working grid program in the reviewed sample is a thin base. Buyers should ask for a measured baseline, an independent evaluation and the energy cost of the model before accepting an emissions claim. The authors also urge customers to choose lower-impact models. Watch which COP commitments name measurable deployments.