Daron Acemoglu

AI does not imply one predetermined future. The crucial question is whether we use it mainly to automate work people already do, or to create new tasks in which human judgment, expertise and responsibility become more valuable. Today, investment incentives push strongly toward automation because firms can readily measure labor-cost savings, while the broader benefits of augmenting workers are harder for any one company to capture. That direction may generate some productivity growth, but it will also concentrate income and decision-making power among large technology companies and owners of capital. Making labor economically dispensable is not merely a distributional problem that can be repaired with transfers. Work helps determine social status, political voice and who participates in important decisions. A society with enormous output but little need for most people’s contributions would be less democratic and less dignified. There is also a knowledge issue. AI recommendations can help people, but if personalized answers replace the effort required to learn and contribute knowledge, short-term convenience may weaken the shared knowledge on which future systems depend. That is a conditional risk, not evidence that knowledge collapse has already occurred. I do not think we should permanently stop innovation, nor do I claim to know an exact AGI timeline. But we should be willing to slow temporarily and redirect development. Public procurement, research funding, worker participation and eliminating tax advantages for capital-intensive automation can encourage pro-worker AI. The future depends less on “AI” in the abstract than on which systems we choose to build, whose capabilities they expand and who gains power from them.

The strongest evidence would be large-scale, sustained adoption of AI that creates genuinely new human tasks—giving workers more expertise, responsibility, bargaining power and higher wages—rather than merely making existing tasks cheaper or eliminating them. If this occurred across ordinary firms, not just in a few celebrated examples, it would make me substantially more optimistic. I would also update if agents overcame today’s application and organizational bottlenecks and produced much larger productivity gains than I expect. They are already impressive enough that I would raise my earlier estimates, though I have not recalculated them. But aggregate productivity alone would not settle the issue. I would still ask who captures the gains and whether workers become more economically and politically dispensable. Conversely, clear evidence that personalized AI is broadly reducing people’s incentives to learn, weakening contributions to shared knowledge, and feeding persistent distortions back into future systems would make me more pessimistic. At present, that is a mechanism from theoretical work, not an established society-wide fact.