Carbon Taxes, Path Dependency, and Directed Technical Change: Evidence from the Auto Industry

GAICIL Insights

Innovation has momentum, and momentum can be a trap.¹ Carbon Taxes, Path Dependency, and Directed Technical Change: Evidence from the Auto Industry, by Philippe Aghion, Antoine Dechezleprêtre, David Hémous, Ralf Martin and John Van Reenen, asks whether directed technical change can help combat climate change. The authors build panel data on auto industry innovation distinguishing dirty patents, for internal combustion engines, from clean patents, for electric, hybrid and hydrogen technologies, using data from 3,423 firms and individuals between 1965 and 2005 across 80 patent offices.

Result one: prices steer innovation. Firms innovate more in clean technologies, and less in dirty ones, when facing higher tax-inclusive fuel prices, a proxy for a carbon tax. Identification exploits differential price movements across countries and firms' heterogeneous exposure through expected sales in different markets. Engineers read price signals as surely as consumers do.

Result two: path dependency. Firms more exposed to clean innovation are more likely to direct future research to clean innovation, a directed knowledge spillover effect. Firms with a history of dirty innovation are more likely to keep focusing on dirty innovation. Path dependency works through aggregate spillovers and through firms' own histories. What a lab knows today shapes what it tries tomorrow.

The implications are stark. Because the stock of dirty innovation exceeds that of clean innovation, path dependency tends to lock economies into high carbon emissions, even after a mild carbon tax or R&D subsidies for clean technologies. This may justify stronger action now, which could be relaxed later as the knowledge stock shifts cleaner. The authors simulate carbon tax increases needed for clean technologies to overtake dirty ones. A gentle nudge loses to a century of accumulated know-how.

The application to AI is direct. Artificial intelligence is subject to the same forces. Researchers and firms choose which directions to pursue, influenced by prices, incentives and past investments. Path dependency means early choices about AI architectures, applications and priorities may shape the field's trajectory for years. Policies influencing those early choices may have lasting effects. The AI field's dominant recipes today may become its lock-in tomorrow.

Timing is the transferable lesson. If path dependency locks in existing directions, acting early to shift incentives beats acting late. Policies steering AI toward socially valuable directions, such as energy efficiency, safety or applications with large social benefits, may be most effective while the field is still developing. The cheapest moment to redirect a technology is before its knowledge stock hardens.

AI can also break the climate lock. AI tools can accelerate research on batteries, materials and energy systems, potentially speeding the shift from dirty to clean technologies. If AI raises the productivity of clean research, it could help overcome the path dependency favoring dirty innovation, reducing the carbon tax needed to tip the balance. AI may lower the price of escaping the past.

AI's own energy appetite connects back. Directed technical change could steer AI toward more energy-efficient methods if energy prices reflect environmental costs. Evidence that prices steer innovation suggests appropriate energy pricing could encourage efficient AI. The same lever that redirected car engineers could redirect model builders.

Incentives are the master key. Carbon pricing, through fuel prices, redirects innovation. Pairing pricing with support for clean research may accelerate the transition. Similar combinations could steer innovation in other domains, including AI, toward desired outcomes.

Pulled together, innovation responds to prices and is path dependent. Higher fuel prices stimulated clean innovation in the auto industry, and past innovation shaped future direction. Because existing knowledge favors dirty technologies, early and strong action may be needed to shift the path. As AI becomes central to the economy, these principles apply to its own development and to its role in accelerating the clean transition.

A hard question for climate and AI policy alike: modest, gradual signals may be too weak against an entrenched knowledge stock. Strong early action, relaxed later, may be both more effective and, over time, cheaper than a long series of gentle nudges.