Companies disclose few uses behind AI climate-benefit forecasts, review finds

Few major companies have publicly announced the specific AI applications underpinning ambitious forecasts of emissions cuts, according to a review published November 9 in MIT Technology Review. Fewer still claim meaningful climate gains from those uses—a gap between projected benefits and documented deployment, rather than proof that AI cannot help reduce emissions.

The findings appear in an opinion article by Mary Nichols, former chair of the California Air Resources Board, and Jeremy Tamanini, founder of sustainability research firm Dual Citizen. Tamanini conducted and reviewed the examination of public information about significant companies and grid operators in power, meat and dairy, and transportation.

The review used Anthropic’s Claude Code, with findings checked against public documents and media accounts. Its test was narrower than whether companies use AI at all: it looked for the emissions-reducing applications assumed in an earlier peer-reviewed study. That distinction matters because general AI adoption does not demonstrate progress along those particular routes.

Published in 2025, that study was led by researchers at the London School of Economics’ Grantham Research Institute and Systemiq. It estimated that AI could help avoid 3.2 billion to 5.4 billion tons of carbon dioxide annually by 2035, after accounting for AI’s own energy footprint. Most of the projected gains came from the three sectors examined in the new review.

Where the projected savings meet deployment

Electricity offers the clearest mechanism. Better AI forecasts of supply and demand could let grids use up to 20% more electricity from existing wind and solar facilities, avoiding an estimated 1.8 billion tons of emissions annually under the earlier study’s assumptions. Yet the review identified only one known, meaningful working program of this kind among the grid operators it examined in China, the United States and Europe.

• Power: A partnership between the UK’s National Grid ESO and Open Climate Fix halved solar-forecast error—the gap between predicted and actual sunlight reaching panels. The article credits the work with avoiding hundreds of thousands of tons of carbon dioxide annually.
• Food: NotCo’s joint venture with Kraft Heinz uses its Giuseppe platform to identify plant combinations that mimic animal products’ taste, texture and smell. The review found few alternative-protein companies with similarly integrated AI development. The forecast depends on closing both taste and cost gaps with conventional proteins.
• Transport: Battery maker CATL says its AI system made screening candidate materials 100 times faster. Separately, the article cites independent researchers who found that Didi Chuxing’s AI-optimized routing cut average per-trip emissions by 16% by reducing detours. The examples show different outcomes: faster development and lower operating emissions.

Better technology still has to reach the market

The authors identify barriers beyond AI performance. Alternative proteins face slow mainstream adoption. More broadly, AI does little to shorten the established processes for certifying, permitting, building, packaging, marketing and selling new products or facilities. Integrating a model is therefore only an initial step toward the changes assumed in the emissions forecasts.

There is also a competing use: oil and gas companies are interested in AI to increase fossil-fuel production. Meanwhile, the article cites an International Energy Agency base forecast in which global data-center emissions could exceed 300 million tons annually in 2035. That figure includes AI and non-AI computing; it is not an AI-only footprint.

Nichols and Tamanini’s prescription is deliberate deployment, not an assumption that benefits will emerge automatically. They urge companies to prioritize emissions-reducing applications, customers to choose lower-impact models and developers, and governments to incentivize or require more responsible operations.