AI and Marginal Revolutions in Wastewater Treatment

A study by French economists, including Nobel laureate Philippe Aghion, found that a predictive machine-learning model for aeration control in French wastewater treatment plants reduced electricity consumption by 5.4%, carbon emissions by 6%, and energy expenditures by 8.2%, while improving effluent quality and resilience to extreme weather. The AI's own electricity demand accounted for less than 1% of the savings, and the authors estimate substantial global welfare gains from CO2 reductions in three diffusion scenarios using the DICE model.

AI and Marginal Revolutions in Wastewater Treatment An interesting paper https://conference.nber.org/conf papers/f240807.pdf from French economists, including recent Nobelist Philippe Aghion, looks at the savings from a predictive machine-learning model applied to wastewater treatment: This paper studies the environmental effects of a specialised AI aeration-control system deployed across French wastewater treatment plants operated by a global leader in water supply services. Exploiting quasi-experimental variation in both the timing of adoption and outages, we estimate the causal impact of AI on electricity use, carbon emissions, and energy expenditures. We find that full-time AI control reduces plants’ electricity consumption and carbon emissions by 5.4% and 6% respectively, and energy expenditures by 8.2%, resulting in negative abatement costs, while also improving water effluent quality. The additional electricity demand generated by AI models represents less than 1% of these savings. Beyond these effects, AI-equipped plants prove more resilient to high operational stress during extreme meteorological events and chemical pollutant peaks. They also improve load management by reallocating electricity consumption from peak to off-peak hours. Finally, we use the DICE model to assess the aggregate implications of our findings in three diffusion scenarios. We find substantial global welfare gains from the CO2 reductions associated with this industrial AI use case. One annoyance: The authors frame the paper as a contrast to worries about AI’s energy use and environmental impact. But those objections are almost entirely innumerate and pretextual and casting the paper as a rebuttal lends them more credibility than they deserve. One note: Don’t misread the “less than 1%” line as being in the same units as the 5.4%, 6%, and 8.2% figures above it — it isn’t a comparable percentage-point offset. It means the AI system’s own electricity draw is a rounding error next to the savings it generates. More generally, the effect of AI will be through many, many improvements of this nature.