Research

Working Papers

1. Chatbots’ Soft Power: Generative Artificial Intelligence Promotes Different Political Values Across Countries

   with Rustam Hakimov and Dominic Rohner. R&R at Nature Human Behaviour

   Abstract PDF

   Generative Artificial Intelligence (AI) has emerged as a transformative technology. Yet, its potential transmission of political values remains underexplored. This study examines whether the most popular chatbot, ChatGPT, adapts its responses to local democratic values, depending on the language in which it is queried and how these responses compare to human opinions across countries from the World Values Survey (WVS). We develop a novel "AI Political Index", quantifying ChatGPT’s alignment with local political attitudes. We find that ChatGPT’s answers partially adapt to local attitudes, but are liberally biased from 35-60%, depending on the question. For specific dimensions of democracy, ChatGPT presents an almost 100% pro-liberal bias. To investigate to what extent this bias may generalize to all generative AI models, we also carry out the identical analysis for one of ChatGPTs greatest competitors, the Chinese chatbot Deepseek. Interestingly, while the qualitative overall picture at first sight is not fundamentally different for Deepseek, we find some salient differences. While Deepseek displays-if anything-even stronger pro-liberal bias for abstract concepts like "democracy", when it comes to specific features of the political system, Deepseek is significantly less weary of having a "strong leader" or of army rule. Last but not least, we also study how the political values displayed by chatbots have evolved over the last six months. While overall chatbot answers have proven very stable over this period, for some questions there have been changes, with notably a salient decline in the commitment to democratic elections and gender equality. The findings of the current contribution offer insights into the normative influence of AI technologies, which may become increasingly salient in the context of global democratic backsliding.
2. Improving Matching Efficiency in Two-Sided Markets: A Mutual Popularity Ranking Approach

   with Tobias Lehmann, Camille Terrier, Rafael Lalive. R&R at JEEA

   Abstract

   Congestion is a common challenge in two-sided markets, yet its costs and benefits remain largely unexplored. Using data from a dating platform, we document significant congestion, characterized by a large excess of demand for certain women. Leveraging exogenous variation in the number of men and women using the dating app, we show that congestion slows down matching time for men while benefiting women by expanding their choice set. This asymmetric effect raises a key question: Can congestion be reduced without negatively impacting women matching outcomes? The second part of the paper explores a natural solution. Men who like a woman who does not like them back generate congestion without improving women choice set. To identify these congestion-inducing interactions, we estimate men and women popularity using eigenvector centrality in the network of likes. We show that likes from men who are significantly less popular than a woman have an almost zero probability of leading to a match, yet they impose congestion costs on other men in the market.

Work in Progress (selected)

1. The Formation of War Support: Economic Trauma and Political Attitudes in Russia

   Work in Progress, slides are available upon request

   Abstract

   This paper studies how exposure to severe economic shocks shapes political attitudes toward violent conflict decades later. Focusing on Russia, I examine whether local economic collapse during the 1990s transition period predicts individual-level support for the Russian invasion of Ukraine in 2022. I combine city-level measures of economic decline based on nighttime lights between 1992 and 2000 with large-scale social media data from VKontakte that capture revealed pro-war attitudes. I find that individuals who grew up in cities that experienced larger economic shocks are significantly more likely to follow pro-war groups today. The effect is concentrated among cohorts who were adolescents or young adults during the crisis. Moreover, exposure to economic trauma interacts with social learning: peer effects strongly predict pro-war behavior, and their influence is amplified among individuals exposed to deeper local crises. The results highlight how long-run economic experiences shape belief formation and political behavior through both individual memory and social transmission.