AI would allow more CO₂ emissions than it avoids, according to global economic modelling

Published in npj Climate Action, a general equilibrium model estimates the net balance of AI at between 0.47 and 1.8 gigatonnes of additional CO₂ per year, due to a lack of sufficient gains from renewables.

A study published in npj Climate Action takes the problem of the climate balance of AI from an unusual angle: not the electricity consumption of data centres, but the effect of AI on the productivity of the energy sectors themselves. The authors model artificial intelligence as a double-way productivity amplifier, integrated into a computable general equilibrium model of the global economy.

The reasoning is simple: the same capabilities that optimize renewable production and efficient use also serve to discover and exploit oil and gas deposits more effectively, extending their economic viability. The authors refer to these additional emissions as "enabled emissions".

A difficult threshold to cross

Of 64 parallel adoption scenarios, the net balance was between 0.47 and 1.8 gigatonne of additional CO2 per year, or 1.2 to 4.8 per cent of global energy emissions in 2024. The most striking result is the identified threshold: once the productivity gains in the fossil sector are non-zero, the emissions allowed outweigh the emissions avoided. To achieve a net decrease, renewable gains would need to be four to five times higher than fossil gains.