Research

While social media has facilitated protests and challenged democratic norms, online information campaigns could also enhance electoral accountability. High saturation campaigns—that target many voters—may ignite social effects beyond direct exposure. We experimentally evaluate the electoral impact of non-partisan Facebook ads in Mexico. Vote shares of municipal incumbent parties that engaged in zero/negligible expenditure irregularities increased by 6-7 percentage points in directly-targeted electoral precincts and 3 percentage points in untargeted precincts within treated municipalities. Both effects were greater where ads targeted 80%—rather than 20%—of the electorate. The amplifying effects of high saturation campaigns reflect interactions between citizens within more socially-connected municipalities.

Social distancing policies have been widely used to curb the spread of infectious diseases such as COVID-19, but assessing their effectiveness is challenging. This study shows that widely-used methods to estimate the effects of such policies are highly sensitive to accounting, or failing to account, for the simultaneous adoption of policies and the presence of spillovers across geographies stemming from human movement. By estimating a series of nonparametric models on fine-grained mobility, epidemiological, and policy data from Mexico, this research shows that failing to consider confounders, interactions, and spillovers can change the magnitude and the sign of estimated policy effects, hampering the design of optimal public policies.

This paper studies the decision of criminal organizations to use violence or bribes to influence political outcomes. Using an asymmetric information model, I show that criminal groups resort to violence when other channels of influence, specifically bribes, become unavailable. I test the model's predictions in the context of an electoral reform in Mexico that increased politicians' cost of accepting bribes. I measure bribes using confidential administrative reports of suspicious financial transactions in retail banking and measure violence using an original dataset of attacks on politicians. Using a difference-in-differences design, I find that in areas with drug trafficking organizations, the reform led to a 4-percentage point decrease (∼650 fewer reports) in suspicious financial transactions and a 2-percentage point increase (∼44 more attacks) in attacks on politicians. Consistent with the model, further evidence indicates that (1) criminal organizations resort to violence when they fail to reach agreements with politicians, and (2) the effects of the reform are driven by municipalities where politicians have tighter campaign spending limits and possess less information about the criminal groups. These findings have implications for our understanding of local governance and call for attention to the design and implementation of transparency-enhancing electoral reforms.

We model how organized criminal groups (OCGs) use bribes, vote buying, and violence to influence elections and officeholders. Post-election bargaining determines the OCG's continuation value from capture, scaling returns to electoral interventions. Our model offers direct predictions: vote buying and voter violence are endogenous substitutes conditional on baseline candidate popularity; OCGs abstain when elections are competitive; pre-electoral violence signals criminal strength; institutional parameters reshape criminal influence non-monotonically. Criminal influence raises turnout and reduces competitiveness when supporting front-runners but depresses turnout and increases competitiveness when supporting underdogs. We discuss the implications for research and policies aimed at countering organized criminal influence.

I examine whether AI-guided reasoning reduces issue polarization in two preregistered experiments conducted on deliberation.io, a purpose-built platform for deliberation research, with 5,000 participants across contentious U.S. policy issues. Relative to reflective writing, AI-emotional regulation, and AI-grammar correction, AI-Socratic dialogue ---which prompts users to articulate supporting arguments--- significantly moderates extreme positions on mental health-based gun regulation. A second experiment comparing AI-Socratic dialogue to AI-grammar correction across abortion, handgun regulation, and voter ID requirements finds that extreme participants moderate their positions and exhibit behavioral changes, including increased cross-partisan donations and reduced endorsement of extreme comments. Moderation effects are largest for abortion, where baseline cross-partisan agreement was highest, and concentrate among participants with low initial confidence, high susceptibility to elite cues, and limited cross-party interactions. Text analysis of conversational transcripts reveals that moderation operates through increased consideration of alternative perspectives rather than improved argument quality, challenging standard deliberative theory. One-month follow-up displays mixed persistence and minimal spillovers to untreated issues, consistent with issue-specific effects. Perceived agency remains equivalent across conditions, suggesting preserved autonomy. Overall, results indicate that AI conversational agents can facilitate compromise through previously underexplored mechanisms, offering promising scalable applications.

Political polarization can weaken electoral accountability by shaping how citizens process information. We examine the impact of disseminating incumbent performance information on voting behavior in a polarized setting and assess the mitigating role of a debiasing nudge. We experimentally evaluate a local CSO's Facebook ad campaign that delivered COVID-19 case and death statistics to over 2 million unique users across 500 Mexican municipalities ahead of the 2021 elections. Polling-station-level results reveal that the information alone backfired: it increased (decreased) incumbent support in areas with high (low) COVID-19 impact. These effects are driven by areas with strong prior incumbent support, prevalent communal values, and higher stress indicators among citizens. However, a debiasing nudge reversed this effect, enabling voters to reward (punish) incumbents with low (high) COVID-19 impact. Our findings underscore how biases in information processing undermine electoral accountability in polarized contexts and demonstrate the potential for nudges to restore it.

We study how large language models (LLMs) can potentially augment survey-based data collected from human subjects, by focusing on two applications: (1) estimating willingness-to-accept (WTA) for giving up digital and analog goods, and (2) predicting personal income levels. We find that supplying LLMs with rich contextual data beyond demographics significantly improves predictive accuracy. Model fine-tuning and retrieval-augmented generation (RAG) further enhance performance, while changes to model temperature or prompting strategies yield only marginal improvements. Performance varies across goods studied and demographic groups. We provide a methodological blueprint for deploying LLMs as a fast, low-cost multiplier of survey coverage. This is particularly relevant in times of rapidly declining survey response rates.

Presented at (1) the 3rd Annual Business & Generative AI Conference, Wharton Human-AI Research; (2) ASSA 2025 Annual Meeting.

Developing public policies that effectively address complex societal issues while representing diverse perspectives remains a significant challenge in governance and policy-making. This paper presents Policy Dreamer, an evolutionary dynamics-based preference aggregation method designed to create public policy that aligns with heterogeneous populations while preserving solution diversity. It does so in three stages: a) Initial Public Policy Generation (where public policies are defined as a set of goals, actions, and strategies aimed at addressing specific societal issues), b) Preference Elicitation from a constituency of humans, and c) Policy Refinement using simulated human feedback. We apply this approach to the domain of creating public policy, which require navigating complex socioeconomic trade-offs. To validate our method, we measure our system's ability to create popular yet diverse policy proposals in the following domains: Healthcare, Gun Control, and Social Media regulation. Our approach iteratively aligns policies with respect to a base constituency, while using evolutionary search to ensure that policy diversity is not compromised. When compared to an expert-crafted set of policies, it is able to generate novel policies, with up to 25% of generated policies being novel. However, it exhibits limitations in capturing the full diversity of these expert-crafted policies, particularly in controversial or emerging policy domains. Overall, our preliminary results suggest that Large Language Models (LLMs) are able to actively elicit preferences from a constituency of people, and iteratively generate statements (public policies) that align with this constituency while preventing a collapse in statement diversity.

Experimental design.

Social media has been identified as a relevant factor in the global rise in political polarization. Yet, the role of political elites in this process remains unclear. Using data on mobile coverage, Facebook connections, and electoral outcomes in Mexico from 2012 to 2021, we investigate the determinants of elite-driven polarization on social media. Our findings indicate that political actors who deepened a divisive discourse during this period gained electoral benefits from increased polarization. We distinguish between mechanisms of selective exposure ("echo chambers") and diversification ("contact"), noting that electoral gains are concentrated where social media amplifies a single political voice. Conversely, exposure to an ideologically diverse network mitigates these effects. By clarifying the role of political elites and differentiating mechanisms of social media exposure, our study reconciles conflicting results in political polarization literature and offers insights into potential mitigation strategies.

Presented at EGAP

We will conduct randomized experiments deploying deliberation.io, an open-source AI-powered deliberation platform, to facilitate constitutional reform discussions in Ghana. Across 150,000 participants, we will estimate causal effects of AI modules on preference formation, cross-group convergence, and opinion quality under high polarization. Results will identify mechanisms enabling collective decision-making and inform cost-effective technologies for participatory development in emerging economies.

Experiment design and platform testing.