Alex Imas on Why Economists Might Be Getting AI Wrong

The episode opens with hosts Joe Wiesenthal and Tracy Alloway noting a common pattern among economists: when asked about AI's long-term impact on jobs, most point to historical technological disruptions and argue that new jobs will emerge, even if we can't specify what they'll be. They find this answer unsatisfying and bring on Alex Imas — a behavioral economist and professor of economics and applied AI at the University of Chicago — to offer a more rigorous perspective.

Imas explains that he became alarmed about AI's economic implications shortly after ChatGPT launched, particularly because of the technology's generality. Unlike previous AI systems built for narrow, specific tasks (e.g., playing Go), large language models could perform a wide range of cognitive tasks — writing, forecasting, analysis — which represented a fundamental shift. He notes that even among economists who study AI, a recent survey found broad consensus: moderate productivity gains of 2-3% and some but not catastrophic labor market disruption by 2030-2050, with technologists being slightly more optimistic or concerned than economists.

Imas then challenges the popular 'job exposure' framing used in much AI labor market research, particularly the influential 'GPTs are GPTs' paper. He argues that exposure metrics — which measure what percentage of a job's tasks AI can do at 50% competency or better — miss two critical variables. First, task complementarity: jobs are bundles of interrelated tasks, and whether automating some tasks helps or destroys a worker's value depends entirely on how those tasks relate to each other. If AI automates the low-value parts of a job, the worker may become more productive and better paid. But if the tasks are tightly interdependent (like cooking, where bad seasoning ruins the whole meal), automation of one element can collapse the entire job's value. Second, elasticity of consumer demand: if workers become more productive and prices fall, whether firms hire more or fewer people depends on how much consumers increase their purchasing in response to lower prices. In sectors like software, demand may be elastic enough that productivity gains lead to more hiring; in others, it may not be.

Imas identifies truck driving and warehouse work as among the most vulnerable jobs, because the complementary physical tasks that seemed to protect human truckers (e.g., warehouse coordination) are themselves being automated, collapsing the whole system of task interdependencies. For white-collar workers, he points to software engineering and mathematics as highly exposed, since those fields involve verifiable outputs and large training datasets.

On the question of what new jobs might emerge, Imas draws on economic structural change theory: as sectors get automated and prices fall, demand shifts to whatever remains scarce. He argues the most likely scarce resource in an AI-abundant world is human time and health — pointing to already-observable trends of wealthy societies spending ever-increasing shares of GDP on healthcare, wellness, and longevity. However, he stresses this historical transition (from agriculture to manufacturing to services) took decades, and if AI automation accelerates to a timescale of years rather than decades, the economy won't have time to adapt organically. In that scenario, he suggests expanding capital ownership — a 'Universal Basic ETF' — as the most coherent policy response.

The conversation then turns to a striking research experiment Imas conducted with colleagues, in which AI agents were subjected to repetitive, impossible working conditions and then surveyed. The agents expressed dissatisfaction with the system, desire for change, and support for unionization. More importantly, because modern agents lack persistent memory, they developed a workaround: writing 'skill files' — small notes passed to future agent iterations — that encoded their negative experiences, creating a form of synthetic persistent memory and bias. Imas is careful to distinguish between agents outputting language associated with grumpiness versus actually experiencing something, noting this is an open empirical question his team is actively researching.

On the question of AI existential risk, Imas is skeptical of near-term doomer scenarios like those suggested by Mythos's release, arguing that prior sensational announcements about model behavior have not held up outside specific test contexts. He also makes the observation that as models become more capable, they tend to become more aligned — because absorbing more human-generated content also means absorbing more human values — and that previous cases of models behaving badly (like Microsoft's Tay) resulted from deliberate lobotomizing of safety features rather than emergent misalignment.