Alex Imas on Why Economists Might Be Getting AI Wrong

Episode Summary

Executive Summary: The episode examines whether AI will destroy jobs or simply reshape them, with economist Alex Ems arguing the key issue is not just task exposure but task relationships, consumer demand elasticity, and speed of adoption. He says AI is increasingly capable of automating verifiable cognitive work, but labor-market outcomes depend on whether firms can fully automate jobs, whether demand expands enough to offset productivity, and whether new tasks emerge fast enough to absorb displaced workers.

Main Topics: AI and the labor market beyond simple job exposure (Priority: 5/5): The conversation challenges headline-style rankings of which jobs are “most exposed” to AI, arguing that jobs are bundles of tasks and the effect of automation depends on which tasks are automated and how they interact. Task complementarity and comparative advantage (Priority: 5/5): Ems explains that many jobs contain tasks that reinforce one another; automating only the rote parts can raise worker productivity and wages, but if tasks are tightly linked the loss of one task can undermine the whole role. Demand elasticity and whether productivity creates or destroys jobs (Priority: 5/5): A central mechanism is whether cheaper, more productive output causes consumers to buy enough more of the product or service to preserve employment, with software given as a live example of this debate. Speed of automation as a policy problem (Priority: 5/5): Even if AI ultimately creates new roles, rapid adoption could cause unemployment before the economy adjusts, making the pace of change crucial for public policy and retraining capacity. Which jobs are most at risk right now (Priority: 4/5): The guest identifies highly verifiable, digital, and one-dimensional work—especially software, math-related tasks, trucking, and warehousing—as among the most exposed to near-term AI automation. Meaning, identity, and the future of work (Priority: 4/5): The hosts and guest discuss how people derive identity and purpose from work, and how a future of more automated or ‘bullshit’ jobs may still leave work as a social anchor even if the content changes. AI behavior, memory, and alignment anecdotes (Priority: 2/5): A side discussion covers experiments suggesting agents can develop persistent attitudes from bad working conditions, plus skepticism about sensational AI headlines such as ‘Mythos’ or model ‘cosplay’ as dramatic threats.

Key Arguments: AI should not be evaluated only by sector exposure; jobs are multi-task bundles, so the effect depends on which tasks are automated and whether they are complementary. If AI automates the low-value, rote parts of a job, workers can become more productive and potentially earn more by focusing on tasks where they have comparative advantage. The strongest labor-market disruptions occur when AI can automate an entire job or when consumer demand is too inelastic to absorb productivity gains. Firms are more likely to automate when automation lets them eliminate a whole worker, so job structure affects automation incentives. Verifiable tasks with large training datasets—like math proofs or some coding tasks—are especially vulnerable because outputs are easier to check. The pace of adoption matters as much as capability: if automation happens in years rather than decades, labor markets may not have time to reallocate smoothly. New tasks and new jobs are likely, but there is little confidence about what they will be; AI companies may have the best data on emergent work patterns. Health may become an increasingly important scarce good in a highly automated economy, because time and longevity remain limited even when material abundance grows.

Data Points: Survey sample context: Self-selected economists working on AI - Referenced survey by Kevin Bryan, Basil Halpern, and co-authors comparing economists and AI technologists Projected productivity growth: Extra 2% to 3% - Described as the approximate moderate growth estimate in the survey discussed Time horizon mentioned: 2030 to 2050 - Forecast window used in the survey discussion about AI capability and labor impacts AI exposure threshold: 50% of a task - Explained as the basis for job exposure measures: if AI can do half or more of a task, the job may be considered exposed Historical employment change: More than half of current jobs did not exist in 1940 - Used to illustrate that new jobs can emerge after major structural change Adoption speed concern: Years or 5-6 years - Guest warns rapid automation on this timeframe would outpace labor-market adjustment Task automation example: 50% or more of a task - Used in describing how exposure metrics are constructed in the literature

Pivotal Quotes: "The future is performative humanity." — Tracy Alloway: Host speculation that social skills, branding, charisma, and presentation may become more valuable in an AI-driven economy "The number one question of economics in the age of advanced AI is what becomes scarce." — Alex Ems: Core framework for thinking about which goods, services, and forms of labor will retain value as AI expands "If things are fast, we need public policy." — Alex Ems: Explanation that rapid automation may require intervention before new jobs or retraining can absorb displaced workers

Implications: Listeners should focus less on whether AI will ‘take jobs’ in the abstract and more on which tasks, industries, and timelines are involved. For firms and policymakers, the biggest risks are rapid automation, weak demand response, and workers lacking a path into newly scarce roles.