From emissions to damage: A mechanism-based assessment of artificial intelligence's environmental cost and the limits of the sustainable energy transition

Artificial intelligence is rapidly transforming the global economy, but its environmental impact presents a clear paradox, as it can improve efficiency while also increasing energy demand. Existing studies focus mainly on emissions, while policymakers care about the associated welfare losses. This paper shifts the analysis from emissions to carbon damage, capturing the monetized cost imposed on economies and societies. It also investigates and quantifies the mediating role of sustainable energy transition in the relationship between artificial intelligence and carbon damage. Using panel data from 37 OECD countries and employing instrumental variable two-step GMM estimation supported by three robustness checks, our results reveal three findings. First, AI significantly increases carbon damage, confirming that AI's scale and energy demand outweigh efficiency gains. Second, AI significantly promotes sustainable energy transition measured through a newly developed composite index capturing renewable deployment, energy access, and efficiency. This reveals the presence of a beneficial indirect channel. Third, the Sobel mediation analysis confirms that this channel offsets a limited share of AI's total carbon damage. Across our estimators, the implied mediation share remains below 15%, indicating that the self-correcting capacity of the sustainable energy transition is real but quantitatively limited. Moreover, a suppression mediation pattern is additionally identified, revealing that studies relying on total effect estimation without mediation decomposition underestimate AI's true direct environmental cost. This finding moves the debate beyond the question of whether AI is good or bad for the environment toward the more policy-relevant question of how much of AI's environmental cost the sustainable energy transition can absorb.