Acemoglu and Gans Debate AI’s Risks, Governance, Inequality and Jobs

Summary

Joshua Gans collects and answers seven questions posed by Daron Acemoglu about the economic, political and social consequences of artificial intelligence. On safety, Acemoglu distinguishes “distorted intelligence” from superintelligence and favors restricting AI to safer domains, while Gans argues that current systems can be both highly capable and distorted because human-set objectives, loss functions and reinforcement-learning judgments shape their behavior. Gans is more open to managing risks after deployment, though he acknowledges hidden biases and unauthorized behavior as serious concerns. On governance, both address the tension between expert control and democratic participation. Acemoglu argues that citizens should help determine AI’s direction, including people outside the United States, Europe and China; Gans favors delegation to experts through elected institutions, with public accountability. On investment and inequality, the exchange examines estimates that AI investment could average 3.6% of GDP annually from 2025 to 2032 and require about $3.7 trillion in annual revenue by 2032 to yield a 10% return. Acemoglu warns that this could sharply increase capital and labor-income inequality, while Gans says the outcome depends on whether revenue comes from labor substitution or productivity growth. They also dispute how quickly AI will improve drug discovery: Acemoglu points to limited downstream evidence and public-health alternatives, while Gans argues that the returns to health-related AI research could still be exceptionally high. On centralization, Gans sees strong competition and open-source alternatives but no clear policy route to decentralize frontier AI. On China, both reject treating AI as a purely zero-sum US-China race and support coordination on safety, although Gans says the countries are often competing in different areas, with the United States focused more on foundation models and China on applications and robotics. On jobs, Acemoglu expects automation to displace more jobs than it creates unless AI enables genuinely new tasks, while Gans is more optimistic about gradual adjustment and says the disagreement mainly reflects different assessments of uncertain future evidence.