I am a PhD candidate in the Economics department of the University of Southern California.
Research
AI Automation and Labor Market Outcomes
Existing measures of occupational AI exposure ignore worker reallocation, failing to capture true economic impact. In this study, I address this by developing a general equilibrium framework that explicitly models the occupational choice problem for workers. Using administrative German data, I recover the unobserved, heterogeneous comparative advantage vectors that drive these choices. I then simulate a productivity-enhancing AI shock derived from task-level automation scores. The results show that the ability to reallocate determines the distribution of welfare gains: generalists capture the largest relative wage gains by pivoting to high-growth sectors. Conversely, specialists, defined by concentrated comparative advantages, experience significantly smaller gains. This relative penalty affects both low skill workers, whose comparative advantage is confined to manual or service tasks, and high-skill workers who are specialized in certain domains, as the limited transferability of their skills restricts them from accessing the highest returns in the labor market.
Work in Progress
Expertise Acquisition and Excessive Automation
With Serap Inci Ozyer Koca. Draft soon.
We show that automation can be excessive when workers cannot finance the acquisition of expertise. In a competitive economy, firms choose whether to automate, and inexperienced workers acquire transferable expertise by performing entry-level tasks. Because workers cannot borrow against future earnings, their pay cannot be negative, and learning must be financed by productive work. When automation stays within the entry-level task range, the remaining work pays for learning and the constraint does not bind. Once automation extends beyond it, firms that do not automate must combine learning with work that juniors perform poorly. At low adoption costs, too many firms automate, experienced workers become scarce, and aggregate consumption falls below the unconstrained benchmark. Along the transition, consumption first rises as juniors shift from learning to production, then falls as expertise runs down. Calibrated to software engineering using task-level usage data from a large language model, stationary consumption falls by 10.5 percent and consumption over the transition by 5.8 percent. We argue that policy should target learning rather than nonadoption: a budget-balanced training subsidy funded by adopters recovers 92 percent of the stationary loss, compared with 40 percent for a rebate to firms that do not automate.
Teaching
Teaching Assistant at the University of Southern California for
- Intermediate Microeconomics
- Intermediate Macroeconomics
