TL;DR

A peer-reviewed study in Nature argues the largest climate effect of AI is not the electricity data centres consume but the fossil fuel that AI helps extract. Applied right across the energy system, the authors put the yearly addition somewhere from 0.5bn to 1.8bn tonnes of CO2 — against a best case of 500mn tonnes saved through AI in renewables.

The asymmetry

Both sides of the ledger are real. Upstream operators use AI to cut drilling costs, locate deposits and lift recovery rates. Solar and wind operators use it to forecast output, schedule maintenance and time battery dispatch.

The problem is the exchange rate between them. On the authors’ modelling, each 1% of productivity won in extraction demands a 4 to 5% gain in renewables just to hold emissions level. “AI-driven productivity gains enable more emissions than they avoid — reinforcing fossil incumbency rather than displacing it,” the report states.

Adoption rates compound it. An IBM industry study found 44% of upstream extractors already use AI when prospecting, with a further 45% intending to within three years. Renewables show far thinner evidence of comparable uptake — and AI cannot clear interconnection queues, planning delays or curtailment, which are the actual constraints.

Scale against the data centre debate

This is the study’s sharpest claim: emissions enabled by energy companies using AI run somewhere between 2.8 and 10 times what the International Energy Agency reckoned data centres emitted last year, a figure it put at 0.18 gigatonnes.

For context on what AI could unlock, Wood Mackenzie estimates AI-assisted extraction could reach 470bn to 1tn additional barrels, while Goldman Sachs projects that in shale it expands reserves by between 8 and 20% while cutting drilling costs by 30%.

Two caveats belong here. The authors are affiliated with a climate advocacy group, the Enabled Emissions Campaign, and the headline range models full adoption rather than describing what has happened.

Looking forward

British debate about AI and climate has fixed almost entirely on data centres — their power draw, their water use, their place in the planning system. Those arguments are running right now across Slough and the Scottish sites, and this batch also carries news of lenders pricing local opposition into project finance.

The study’s contribution is to point out that the meter everyone is watching may be the smaller one. Its policy recommendation — count enabled emissions, not just direct consumption — would change what any AI sustainability disclosure is actually measuring.