AI isn't just using energy. It's helping fossil fuel companies find more.

A peer-reviewed paper in npj Climate Action proposes a climate-accounting category the tech industry does not currently use. The authors call it "enabled emissions": the pollution AI tools generate indirectly by helping oil and gas companies find, develop, and extract more fossil fuels. Their estimate puts that downstream footprint somewhere between Mexico's annual emissions and Russia's.

The data center power draw that dominates AI climate coverage is only one side of the math. The new paper argues the larger, indirect line is the one nobody is counting.

The paper was written by Will and Alpine. They left Microsoft in early 2024 over the company's continued work with oil and gas clients. The Alpines turned that perspective into a research project: build the missing accounting category, name it, and put a number on it.

Their low-end estimate is that AI-enabled fossil fuel production could add emissions equal to Mexico's annual output. The high end, which assumes more aggressive industry adoption, is on par with Russia's, the world's fourth-largest national emitter. Either way, the figure outpaces the emissions AI is expected to add through its own data center buildout. "It is two sides of the same coin," Will Alpine told Wired via Grist's Climate Desk. "You cannot have artificial intelligence without the energy system that powers it, and that energy system is still dominated by fossil fuels."

The range is wide on purpose. The paper lays out high and low scenarios rather than a single point estimate, in part because the underlying assumption is how aggressively oil and gas producers adopt AI across exploration, drilling, and reservoir management. That adoption curve is the part most exposed to industry pushback, and the part most worth watching as the paper circulates.

The reason the gap exists is structural. Tech companies report their own operational footprint (the electricity their data centers pull) and their supply-chain footprint (the hardware and buildings that go into them). What they do not report is the pollution their customers create with the help of those tools. Oil and gas companies have used machine learning to interpret seismic surveys, optimize drilling, and manage reservoirs for years. Each of those workflows can pull more barrels or cubic feet of gas out of the ground per unit of effort, and each marginal barrel carries its own combustion emissions somewhere downstream.

This is the supply-demand loop the Alpines describe. AI gets cheaper and more capable, fossil fuel producers apply it to their operations, more fossil fuel gets produced, the energy system stays carbon-heavy, and AI's data centers stay powered by it. The two reinforce each other. A framework that only counts the data center side misses the larger line, which is the point of naming it.

The paper also takes aim at a common counterargument: that AI accelerates clean energy too, by improving solar forecasting, grid balancing, and materials research. The Alpines' reading of the literature is that the fossil-fuel side of the ledger grows faster. They argue the net effect is upward, not downward, and that the gap between the two sides is the part current reporting cannot see.

The framework is now on the table. Cities, regulators, and reporters can ask for an "enabled emissions" line the way they already ask for scope 1, 2, and 3 disclosures. The paper does not prescribe how to allocate that figure across AI providers, but it gives the conversation a name, a range, and a unit of measure it did not have before. The next test is whether major AI customers and their shareholders treat it as a category they can be asked about.