A new AI climate model puts cheaper fossil supply at the center of the risk

Those are not symmetric economic interventions. The model finds that productivity in upstream fossil extraction is the main source of the gap: lower marginal costs can make marginal projects viable, increase available supply and feed through to consumption and other sectors. The paper labels the resulting additional CO₂ "enabled emissions"; it calls reductions associated with optimization "avoided emissions."

Under uniform productivity shocks, the study estimates that lower-carbon generation gains would need to be roughly four to five times fossil-sector gains to reach a modeled CO₂ breakeven. Across its 64 adoption combinations, emissions fell only when fossil productivity gains were set to zero. That finding applies to this model’s structure and parameter choices, not to every application marketed as AI for energy.

The adoption benchmark is useful—and easy to overread

“Parallel adoption” does not mean the model assumes identical productivity gains. It means each pathway is set to the same adoption tier while retaining different, evidence-based gain ranges. The paper says it used 0–20% gains for fossil pathways, 2–30% for lower-carbon generation, 2–10% for grid efficiency and 2–15% or 2–6% for the two modeled demand-side applications.

That choice cuts in two directions. The authors treat the headline benchmark as conservative because evidence of fossil-sector deployment is more mature, while many lower-carbon applications are pilots or estimates of technical potential. The study also parameterizes lower-carbon generation at the upper bounds of that potential. Yet a parallel-tier scenario still cannot establish the real-world pace, mix or geography of adoption.

There is evidence that the supply-side channel is already more than hypothetical. Producers including Chevron, ExxonMobil, ADNOC and Aramco say they are using AI to identify prospects or improve recovery, while SLB, Halliburton and Baker Hughes are applying it to drilling and well placement, according to a contemporaneous account. Saudi Aramco has said it embedded AI "in everything," while Equinor credited seismic technology and AI for discoveries; those are company claims, not independent measurements of AI’s net climate impact.

The oil industry disputes the study’s framing. Andrea Woods, a spokesperson for the American Petroleum Institute, said the U.S. industry is producing more energy while reducing emissions through technology, operational practices and policy. That response speaks to emissions intensity, while the paper’s modeled question is total CO₂ after economy-wide supply and demand responses. The two claims therefore do not directly settle each other.

The study’s diagram of direct, indirect and system-level AI–energy-market effects, including modeled, partly modeled and unmodeled mechanisms. Source: npj Climate Action.

What the estimate leaves outside its boundary

The 0.47–1.8-gigatonne range covers CO₂ from fossil-fuel combustion. It excludes methane, other non-CO₂ gases and industrial-process emissions. The authors say those omitted non-CO₂ effects are not directionally determined: AI could reduce methane through leak detection while expanded extraction and industry could increase emissions.

It also excludes first-order data-centre emissions from the equilibrium calculation. The paper compares its modeled fossil-enabled emissions, 0.6–2.4 gigatonnes annually in isolated pathway cases, with an International Energy Agency estimate of 0.18 gigatonnes for data centres in 2025. That is context, not an addition: the study explicitly says the two analytical frameworks should not be summed.

Other constraints matter just as much. The model uses a 2017 GTAP-Power database, adjusted so the electricity mix approximates 2024 levels, and aggregates the world into five regions and 20 sectors. It is static, assumes full employment and capital reallocation, and cannot resolve particular firms, projects, changing adoption rates, grid physics, geological limits or future infrastructure lock-in. Nor does it represent green hydrogen, small modular reactors or direct air capture; the authors say their exclusion understates possible avoided emissions but do not quantify the effect.

The code, configurations and output data are available in the authors’ repository, although reproducing the full workflow requires Windows-compatible GEMPACK software and the underlying GTAP material has separate terms. That transparency makes the assumptions easier to test; it does not turn the model into an observed causal estimate.

The decision is whether deployment evidence can replace the scenario

The study supports a narrower conclusion than “AI is bad for the climate”: productivity gains directed at a fossil-heavy economy can raise total CO₂ even when the technology improves lower-carbon operations. It also supports a policy argument from its authors—that AI climate assessments should track enabled production as well as operational power use and avoided emissions. Independent reporting on the paper similarly highlights the difference between improving renewable operations and proving an economy-wide reduction in emissions.

The next evidence needed is less another technical-potential claim than measured deployment: which AI systems change extraction costs, production volumes, curtailment, grid capacity or end-use demand, in which markets, and with what counterfactual policy constraints. A dynamic model built on those observations could test whether the paper’s structural asymmetry persists as energy systems, infrastructure and regulation change.