A working paper by French economists, including recent Nobel laureate Philippe Aghion, examines the savings generated by a predictive machine-learning model applied to wastewater treatment. The study analyzes a specialised AI aeration-control system deployed across French wastewater treatment plants operated by a global leader in water supply services.

Aeration—the process of injecting air into wastewater to sustain the microbes that break down organic matter—is typically the single largest electricity consumer at activated-sludge treatment plants, frequently accounting for 40–60% of a facility's total energy use. Optimising this step has long been a target for efficiency improvements, but conventional control systems rely on fixed schedules or manual adjustments that cannot respond in real time to fluctuating inflow volumes, pollutant loads, and weather conditions.

Using quasi-experimental variation in both the timing of adoption and system outages, the researchers estimate the causal impact of AI on electricity use, carbon emissions, and energy expenditures. According to the paper, full-time AI control reduces plants' electricity consumption by 5.4% and carbon emissions by 6%, while cutting energy expenditures by 8.2%. These reductions result in what the authors describe as negative abatement costs, meaning the system pays for itself in energy savings while simultaneously improving water effluent quality.

The paper notes that the additional electricity demand generated by the AI models themselves amounts to less than 1% of the savings achieved. Importantly, that figure is not a directly comparable percentage-point offset against the 5.4%, 6%, and 8.2% reductions cited above; rather, it indicates that the AI system's own electricity draw is negligible relative to the savings it produces.

Beyond direct efficiency gains, the study finds that AI-equipped plants prove more resilient to high operational stress during extreme meteorological events and chemical pollutant peaks. The systems also improve load management by shifting electricity consumption from peak to off-peak hours—a capability that aligns with grid operators' growing need for demand-side flexibility as intermittent renewables take a larger share of electricity generation.

To assess the broader implications, the authors apply the DICE (Dynamic Integrated Climate-Economy) model—the integrated assessment framework developed by 2018 Nobel laureate William Nordhaus that links economic activity to climate outcomes—across three diffusion scenarios. They conclude that the CO2 reductions associated with this industrial AI use case yield substantial global welfare gains.

The full paper is available from the National Bureau of Economic Research (NBER).

Commenting on the paper at Marginal Revolution, the blog's author notes that the researchers frame their findings as a contrast to concerns about AI's energy use and environmental impact. The author characterizes those objections as "almost entirely innumerate and pretextual" and argues that casting the paper as a rebuttal lends them more credibility than they deserve. The author further observes that the broader effect of AI will come through many incremental improvements of this nature across a wide range of industries.