More Efficient Oil Is Still More Oil

Research published in the journal npj Climate Action estimates that artificial intelligence could increase global energy emissions by up to nearly 5 percent by making fossil-fuel companies more productive. As reported by WIRED, the projected increase would substantially exceed estimates for the direct emissions of the global data-center buildout and could outweigh AI’s modeled benefits for cleaner energy technologies.

The paper was written by Will Alpine and Holly Alpine, former Microsoft sustainability workers who left the company in 2024 over its work with the oil and gas industry. Their argument shifts attention from the electricity used to train and run software toward what customers can produce with it. Oil and gas companies already use computational tools to locate deposits, develop fields and improve operations. Better software can make those activities cheaper or more productive.

The missing line on the invoice

Efficiency sounds like conservation because, in a household, using less fuel for the same trip usually is an improvement. A producer faces a different incentive. If AI lowers the cost of finding or extracting a barrel, projects that once looked marginal may become profitable, existing operations may yield more output, and additional supply may encourage additional consumption. The producer saves money; the carbon ledger can grow.

The researchers call the resulting pollution “enabled emissions.” That category differs from operational emissions, which include the power used by data centers and the emissions in a technology company’s supply chain. Operational accounting asks what the software consumes. Enabled-emissions accounting asks what the software helps its buyer sell. Both matter, but they answer different questions.

The nearly 5 percent figure remains an estimate, not a measured outcome. It depends on assumptions about productivity, fossil-fuel supply, demand and how markets respond to lower costs. AI might also improve maintenance, reduce leaks or support renewable-energy development. Those uses deserve credit where total emissions actually fall, but they do not cancel increased extraction by rhetorical addition. A solar forecast and a more profitable oil field remain separate consequences.

That distinction gives buyers and regulators a practical test. Before calling an AI deployment efficient or sustainable, count the carbon associated with the activity it expands, not just the servers it occupies. Ask whether total emissions decline against a credible baseline, who verifies the result and whether lower operating costs enable more production. Otherwise the climate claim records the discount and discards the purchase.