New working papers
1. Strategic Human-AI Interaction and (mis)Alignment, with Gilat Levy.
Abstract: We explore the implications of human-AI interaction when the human is strategic and is aware that her decisions affect the training of the AI agent. The human employs an AI agent that may act on her behalf to save on cost. While typically AI agents learn from human behaviour using the assumption of revealed preferences, we focus on environments in which choices do not necessarily reveal preferences and hence some misalignment may arise. We show that strategic training implies that AI agents can sometimes allow the decision maker to attain higher welfare than the one that arises from just lower decision-making cost. This arises as misalignment can sometimes be beneficial, and moreover, misalignment can be induced strategically by the decision maker to her benefit.
2. The Occasional Voter Problem: Preference Aggregation in Multiple Elections and AI-Assisted Voting, with Gilat Levy
Abstract: We study preference aggregation when voters face multiple simultaneous elections and participation is occasional --- that is, voters can engage in only a subset of decisions. A key feature of the multi-issue environment is that preferences across elections are positively correlated. This correlation, combined with the independence introduced by occasional participation, generates a fundamental inefficiency: for sufficiently high correlation the voting mechanism does no better --- and can do strictly worse --- than an uninformed educated guess. We call this the Occasional Voter Problem. We characterise this inefficiency precisely using multivariate normal orthant probabilities, and show that occasional voters who participate in both elections impose a negative externality on representation in the election they do not primarily attend. We then analyse AI-assisted voting, in which an AI agent trained on historical data complements single votes across elections. Under exogenous participation, AI assistance weakly improves representation when preference correlation is sufficiently high. Under endogenous costly voting, however, AI assistance creates differential incentives across voter types: aligned voters (whose preferences co-move across issues) benefit disproportionately, generating a representation bias in favour of the majority. We characterise equilibrium voting strategies, quantify the trade-off between higher turnout and biased representation, and identify conditions under which AI assistance improves aggregate welfare.
3. Research Waves, with Mariagiovanna Baccara and Ronny Razin (CEPR discussion paper DP19808, CESifo working paper 12248). For additional proofs not in the paper's main Appendix, see the Supplemental Appendix.
Abstract: Competing lines of research start and grow as information is gradually uncovered by scientists who, driven by career incentives, sort themselves into fields. We build a strategic experimentation framework in which agents irreversibly specialize in one of two risky fields, and information updates are more frequent as more agents specialize in a field, and we describe the equilibrium forces that determine the size, shape, and length of such research waves. We show that in the bad news case, all researchers specialize in one field, generating a unique bandwagon wave. As the difference in prior beliefs increases, such wave starts earlier, grows more slowly, and becomes longer. In the good news case, both fields can be explored in equilibrium in two sequential surges. The probability of both fields being explored increases in the researchers' pool size and in the efficacy of technology. We also assess the impact of policies used in the scientific community, such as
citations benefits, tenure clocks, and grants, on the structure of the equilibrium research waves.
Paper available upon request