On the Carbon Footprint of Economic Research in the Age of Generative AI

A new study titled 'On the Carbon Footprint of Economic Research in the Age of Generative AI' shifts the environmental conversation from AI models to the workflows they enable. Researchers Andres Alonso-Robisco, Carlos Esparcia, and Francisco Jareño analyzed how generative AI tools like GitHub Copilot and ChatGPT are expanding computational workflows in economic research, creating new environmental impacts beyond just model training. They mapped the Green AI literature into seven themes, finding that while training footprint remains the largest concern, inference efficiency and system-level optimization are rapidly growing research areas alongside measurement protocols and governance frameworks.

The researchers benchmarked a modern economic survey workflow—specifically an LDA-based literature mapping implemented with GenAI-assisted coding in a fixed cloud notebook environment. Using CodeCarbon to measure runtime and estimated CO2e emissions, they discovered that generic 'green' language injected into prompts had no reliable environmental benefit. However, implementing specific operational constraints and decision rules in prompts delivered substantial and stable footprint reductions while preserving decision-equivalent topic outputs. This suggests that human-in-the-loop governance, where researchers strategically allocate discretion between themselves and AI systems, represents a practical lever for aligning GenAI productivity with environmental efficiency.

The study's methodology treats prompts as decision policies that govern what code gets executed and when iteration stops, offering a novel framework for measuring and optimizing the environmental impact of AI-assisted research. By focusing on the workflow level rather than just the model level, the research provides actionable insights for researchers and organizations looking to reduce their computational carbon footprint without sacrificing research quality or productivity.