Date of Award

2026

Degree Type

Open Access Dissertation

Degree Name

Economics, PhD

Program

School of Social Science, Politics, and Evaluation

Advisor/Supervisor/Committee Chair

Gregory DeAngelo

Dissertation or Thesis Committee Member

C. M´onica Capra

Dissertation or Thesis Committee Member

Scott Cunningham

Terms of Use & License Information

Terms of Use for work posted in Scholarship@Claremont.

Rights Information

© 2026 William Wyatt

Keywords

AI, Bail, Behavioral Economics, COVID, Game Theory, LLM

Subject Categories

Computer Sciences | Criminology | Economics

Abstract

This dissertation contains three studies. Each asks how rules or language change the choices people and machines make when outcomes are uncertain. The first study, written with Kiran John, evaluates California’s 2020 cashless bail reform. We use propensity score matching on arrestee records from the windows before and after implementation, and we test whether the shift away from cash bail produced any effect on subsequent offending. It did not. Matched comparisons yield small, statistically insignificant differences across every window we examined. That null result cuts against both sides of the public argument. The reform did not drive a spike in crime, and it did not measurably improve outcomes for released defendants either. The honest answer is closer to “nothing happened,” at least in the data we have. The second study, with Gregory DeAngelo and Bryan C. McCannon, runs a Stag-Hare coordination game with large language models as players. Stag-Hare is the cleaner variant of Stag Hunt: two players each pick the safe option or the cooperative option, and the cooperative payoff dominates only when both pick it. In humans, prompting players to think about responsibility for the outcome tends to reduce risk-taking. We find the opposite pattern in models. When a language model is told that it bears responsibility for the consequences of its choice, it cooperates more often and picks the riskier stag more often, not less. We document the reversal across several model families and prompt variants and discuss what this suggests about how responsibility framing maps onto model behavior, which turns out to be almost backward from the way it maps onto human behavior. The third study proposes an embodiment-scoring framework for evaluating LLM responses to behavioral protocols and asks whether prompt framing changes decisions directly or changes the reasoning the model produces on the way to a decision. Using decision tasks drawn from the experimental economics literature, I score model outputs on how much first-person perspective they adopt and whether their stated reasoning invokes participation or observation. Prompt framing reliably moves the reasoning mode. It does little to the final decision. For researchers using LLMs as synthetic human participants, this matters. The surface behavior can look stable while the internal narrative the model is generating to arrive at that behavior shifts with the prompt, and that narrative is often the object of study in the first place. The three papers share a methodological point. Bail reform, coordination games, and prompt design each involve a visible decision that hides a lot of machinery. When researchers study only the decision, they miss where the interesting variation actually lives.

ISBN

9798244858822

Download

Share

COinS