Working papers
AI is growing at a rapid pace and it is nowhere more evident than in the growth of data centers. Leveraging a novel dataset of U.S. data center energy loads, utility prices, and establishment-level outcomes, we quantify the local spillover effects of the AI revolution on electricity prices, firm performance and labor decisions. Surprisingly, we find no local spillovers over 2010–2024, although the recent period displays a noisy increase. A regional model calibrated to the empirical null suggests that shocks larger than those observed through 2024 will result in noticeable increases in household utility bills if not offset by regulation or external supply.
Do social media platforms enforce their rules uniformly? Evidence from Suspensions on Twitter
I examine Twitter's account suspension decisions following the 2020 U.S.\ election, asking whether verified users enjoyed preferential treatment. Leveraging the \textit{VoterFraud2020} dataset, which covers over 50{,}000 influential accounts, and exploiting Twitter’s abrupt enforcement shift after January~6, 2021, I estimate the causal effect of verification status. Using instrumental variable and control function methods, I find that verified users were between 5 and 15 percentage points more likely to remain active, even after controlling for tweet toxicity, political network affiliation, and account characteristics. While focused on a politically charged moment, the findings highlight broader dynamics in platform governance, where user status and strategic incentives shape enforcement decisions.
Non-competition Agreements and Dedicated Human Capital (Job Market Paper) Github, Google drive
How does the optimal stringency of a non-competition agreement (noncompete) vary with the employee's position inside a firm's hierarchy? I propose a theoretical model in which the employee's productivity increases with their position. A noncompete tilts the holdup power towards the firm. In equilibrium, employees in top positions are subject to a noncompete and the firm promises high compensation to ensure that they exert effort. Employees in middle positions are free from the covenant so that they maintain their incentives to exert effort for a lower wage. Strikingly, noncompete reappears at the bottom of the firm's hierarchy. Since the employees' productivity is low, their compensation does not incentivize effort. A policy to ban noncompetes for bottom positions increases social welfare if the training the firm provides is sufficiently valuable outside the firm and the firm dismisses employees infrequently.
The Economics of Non-competition Clauses
A crucial organizational decision is the degree of access an employee is provided to the critical asset of the firm. Access increases the employee's productivity inside the firm, but also enables the agent to compete upon leaving the firm. I study the optimal compensation package of an employee in terms of access, wage, and noncompete for different levels of ability. I show that the firm compensates the lowest ability agents via access while offering the minimum wage and strictest noncompete, since access not only increases the employee's utility but also the firm's production. For higher ability agents, the maximum degree of access is provided, while the rest of the compensation depends on the damage the employee causes with competing. With low damages, the firm offers a lax noncompete with smaller wages. If the damage is high, higher wages and stricter noncompete are offered. The pattern based on damages is observed in CEO contracts [Kini et al., 2020].
How to enforce platforms' liability?
The role of online platforms has grown substantially in recent years, leading to an increase in illegal content uploads. New regulations are being developed to tackle the issue, such as the Digital Services Act by the European Union. This paper develops a model to illustrate that policing on online platforms is not only a technological issue but a matter of economic incentives, too. In the model, heterogeneous agents of a strictly liable platform may violate the rules set by a policymaker. The higher the type of the agent, the larger the benefit she brings to the platform, but also the larger the societal harm if she violates the rules. The paper argues that overly simplistic regulations can actually worsen the problem of "cherry-picking", where platforms punish only low-type agents and not high-type ones. To avoid cherry-picking, the optimal sanction may be lower when more agents commit a violation. Based on Twitter unsuspension data, I provide evidence consistent with cherry-picking.
Work in Progress
A Theory of Employment Based on Intangible Capital
Optimal Liability for User Generated Content