Working Papers

Recommendation algorithms shape not only how content is distributed but also what producers choose to supply. This paper provides empirical evidence of such supply responses and quantifies their implications for platform design. 

Combining matched print and online headlines from major newspapers with over 3 million Facebook posts, I document that online headlines are more ideologically slanted than their print counterparts. Exploiting Facebook’s 2025 restoration of political content, I find that the online–print slant gap for political relative to non-political stories increases by about 81%, revealing a substantial supply response to the change in algorithmic incentives.

I then develop a structural model of user demand, algorithmic recommendations, and editorial choices. The model shows that media outlets adjust slant primarily to attract algorithmic exposure, with supply decisions accounting for approximately 80% of extra slant in viewed content. Counterfactuals show that removing algorithmic amplification eliminates roughly 90% of this supply-induced slant. 

These findings show that evaluating platform algorithms requires accounting for their effects on content supply, rather than treating the available content as fixed.

This paper studies how an information provider designs disclosure to keep users engaged when users would prefer to learn quickly. The framework applies to expert services, online content, and personalized news environments. We show that optimal disclosure depends on relative patience, belief disagreement, and whether communication can be personalized. When beliefs differ, the principal may cater to the user’s bias by initially revealing information aligned with that bias. Comparing personalized and non-personalized communication, we find a trade-off: non-personalized communication delivers information faster but with lower quality, while personalization slows disclosure but can improve learning. The paper provides a theory of engagement, personalization, and information quality.

Automated market makers are digital trading platforms that replace order books with algorithmic pricing rules. This paper studies how these rules shape price discovery when liquidity providers respond strategically to trades. We develop a dynamic model of liquidity provision under adverse selection and test its predictions using 19.2 million transactions from 31 Uniswap v2 pools. We show that liquidity providers are not purely passive: many actively adjust prices through swaps, and their trades tend to offset prior liquidity-taker trades, especially when those trades appear less informed. The results highlight how automated platform design shapes strategic behavior and market outcomes.