AI is making oil and gas cheaper to find. A new study says that outweighs its clean-energy gains.

The way to make AI good for the climate runs through whatever industry is already wired to absorb it. A new peer-reviewed study modeling 64 ways AI could reshape the power sector finds that, in nearly every scenario, AI's net effect is to raise annual emissions rather than lower them, because fossil-fuel operators are already absorbing AI's productivity gains while clean-power deployment is still stuck at the pilot stage.

The researchers at the University of Michigan and Princeton, writing in Nature's npj Climate Action, report that across their scenarios AI would add between 0.47 and 1.8 gigatonnes of CO₂ to the atmosphere each year. That range is roughly 1 to 5 percent of the energy sector's current annual emissions, and it does not include the electricity used by the data centers themselves. The study is the first to model AI's impact across the full power sector at once, rather than pitting data-center demand against emissions savings from AI-optimized renewables, a comparison the authors call inadequate.

The asymmetry shows up in the math. For AI's emissions impact to break even at equal adoption rates between clean and dirty energy, AI-driven productivity gains on the renewable side would have to outpace fossil-fuel productivity gains by at least a factor of four. The model finds net emissions only fall in scenarios where AI does not raise fossil-fuel productivity at all. In scenarios where AI does both dirty work and clean work, the fossil side runs away with it.

The deployment gap is what makes that 4× ratio hard to clear. On the fossil side, the work is already wired in. Saudi Aramco said last year it had embedded AI "in everything," from seismic imaging to drilling optimization, and credited the technology with production gains. The IEA estimates that AI could eventually boost technically recoverable oil and gas reserves by 5 percent and cut costs at deepwater offshore projects by 10 percent. Industry framing has been blunter. Some executives have called the current AI rollout in upstream oil and gas "the next fracking boom", shorthand for the early-2010s surge in shale oil and gas production that reshaped global energy markets.

On the renewable side, AI is still mostly being tested. Solar and wind forecasting, grid balancing, and predictive maintenance for turbines are active research areas, but most utility-scale deployments are pilots. The bottlenecks are not model quality. They are permitting, interconnection queues, and the slower-moving contracts that govern how new generation gets built and paid for in regulated markets. AI cannot accelerate a transmission-line review.

Lynn Kaack, an assistant professor of computer science and public policy at the Hertie School in Berlin who reviewed a draft of the paper, said prior research on AI's climate effect "completely omit[s] this picture of AI causing increases in emissions" by focusing only on the demand side. Her point, and the paper's, is not that AI is uniquely harmful. The same productivity tools, applied at scale to solar, wind, and grid operations, would help. The question is which industry is actually positioned to use them first, and the answer right now is oil and gas.

Fossil-fuel operators have decades of in-house data, mature digital twins, and procurement budgets for AI tooling. Renewables are fragmented across thousands of smaller developers, municipal utilities, and ISO market rules, with thinner data histories and tighter capital constraints. The 4× breakeven the authors identify is, in practice, a deployment problem dressed up as a model-performance problem.

Co-author Holly Alpine co-founded the Enabled Emissions campaign, a research and advocacy group that tracks AI use in fossil expansion. Her advocacy role is part of why the paper landed in The Guardian as a quantified critique rather than a wire-style summary, and it is part of why an independent academic reader was sought out. The numbers themselves are modeled, not measured: they come from a set of 64 explicit scenarios the authors built, each combining different assumptions about adoption speed, productivity gains, and grid mix.

Standard AI-and-climate coverage reads this as a data-center problem. The study's authors reject that reading. The same AI tools can be a climate win, but only if the industries that can deploy them fastest are not also the ones with the largest existing carbon footprint. The same productivity bump that makes a deepwater oil field cheaper to run can, in principle, make a solar farm cheaper to build and a wind farm easier to interconnect. Which outcome shows up first depends on which side of the energy economy has the workflows, contracts, and capital to actually run the tool.

The testable question, then, is whether renewable-side AI deployment can outrun fossil-side AI deployment. Interconnection reform in the United States, faster permitting in Europe, and the maturation of grid-forming inverters (the power electronics that let wind and solar plants stabilize the grid on their own) are all on the table. None of them is a model problem. All of them are policy problems. Until renewable-side AI moves at the speed the math requires, the model says AI is, on net, a climate loss.