Machine learning may harm the climate more than data centers: new study warns

npj Climate Action has published a study making a broader case for how we judge machine-learning systems: not just by the electricity they use, the water they draw, or the construction needed to run them, but by the pollution they help unlock in fossil fuel extraction.

The paper estimates that the productivity gains AI brings to oil and gas could drive emissions 3 to 13 times higher than the direct emissions from data centers themselves. In other words, cleaner chips, better cooling, and more efficient servers may not count for much if those same tools make it easier to pull more coal, oil, and gas out of the ground.

The authors argue this isn’t hypothetical. They point to Shell, ExxonMobil in Guyana, Chevron, Saudi Aramco, Petrobras, and companies in China already using AI for seismic work, subsurface analysis, exploration, and production. Their argument is that even if AI helps with grids, forecasting, and renewables, those gains may be outweighed because fossil fuel uses are already mature, profitable, and ready to scale.

The study also brings cloud companies into the picture. The authors note that Microsoft and Amazon Web Services have both drawn criticism for their ties to oil, that Microsoft signed a 20-year deal with Chevron linked to a gas-powered data center, that Google says it won’t build upstream systems, that Microsoft’s emissions rose 25%, and that Amazon’s Texas site can emit 33 million tons of CO2 a year. They also say the usual discussion still misses a lot outside the U.S. and beyond large multinationals, including governments, local communities, and expansion across Africa and Latin America.

If you track AI’s climate impact, this one’s worth reading in the journal.