Date of Award

Spring 2023

Degree Type

Open Access Dissertation

Degree Name

Economics, PhD


School of Social Science, Politics, and Evaluation

Advisor/Supervisor/Committee Chair

C. Mónica Capra

Dissertation or Thesis Committee Member

Joshua Tasoff

Dissertation or Thesis Committee Member

Thomas Kniesner

Terms of Use & License Information

Terms of Use for work posted in Scholarship@Claremont.

Rights Information

© 2023 Shanshan Zhang


Deception, Experiments, Gender, Voice Analysis, Voicebot

Subject Categories



The U.S. economy loses hundreds of millions of dollars in tax revenues, wages, and investment dollars, as well as hundreds of thousands of jobs each year due to dishonest behavior (Mazar and Ariely, 2006; Griffin et al., 2022). Thus, understanding dishonest behavior and finding mechanism to effectively reduce dishonest behavior are of great relevance to policy makers and the economy in general. This dissertation studies deceptive behavior and the relevant factors using online experiment, observational data, and field experiment in three chapters, respectively. Chapter 1 studies the effect of interaction with a machine (voicebot) on dishonest reporting. We conducted an online experiment using a coin-toss task and compared reported outcomes across different reporting channels: Human Voice, Voicebot, and Text. We designed a uniform online voice chat interface to standardize the reporting experience. We also tested the effect of a feminine and a masculine voice on misreporting and varied the level of sophistication of the voicebot (AI-enhanced voicebot). Our results show that, on average, there is no significant difference in the likelihood of misreporting through a voicebot and a human voice, or between verbal and written reporting. However, we found that participants who listened to a feminine voice were more likely to lie than those who listened to a masculine voice. Moreover, those who heard a feminine voice were more likely to lie to a voicebot than a human voice. Interestingly, such difference disappears with higher sophistication (i.e., AI-enhanced voicebot). In contrast, when hearing a masculine voice, there was no difference in misreporting between the voicebot and human voice treatments. These findings suggest that utilizing a masculine voice for voicebots or voicebots with higher sophistication and feminine voice could help deter or diminish dishonest reporting in human-machine interactions. Chapter 2 focuses on lying detection from voice analysis. We use video clips from the British TV game show “Golden Balls”, where the contestants play a prisoner’s dilemma game with pre-play communication and from “Real-life Trial” (RT) data, which consist of videos collected from public court trials. We first apply machine learning model to predict the cooperative and deceptive behaviors from acoustic features. We then identify which acoustic features are associated with cooperation or deception. Our machine learning models achieve an average prediction accuracy from 58% to 79%. This suggests that acoustic features are effective in predicting cooperation and deception. We also find that the pitch (fundamental frequency) is positively associated with both cooperation and deception in different contexts. The intonation (the standard deviation of pitch) is negatively associated with deception. In chapter 3, we examine dishonest behavior using a field experiment. We hypothesize that clothes can affect the behavior of the wearer by influencing the person’s identity. We test this hypothesis by recruiting trick-or-treaters during Halloween, a time of year when people wear salient and extreme clothing. We use the lying game of Fischbacher and Föllmi–Heusi as our experimental paradigm with 2 × 3 × 2 conditions. First, we vary the stakes to price lying behavior. Second, we run three conditions with different beneficiaries of the report (self, other, and both) to test whether lying for others is perceived to be normative. Third, we manipulate the salience of one’s costume to test the effect of costume and identity on ethical behavior. Surprisingly, we find that costume salience caused “good guys” to lie more and “bad guys” to lie less. We interpret this either as a moral licensing effect or as stemming from a perception of being monitored. Our design allows for the identification of contagion effects, and although there were no direct effects of gender, we find that children lie more when children of the same gender near them lie more. We also find that stakes had no effect, people lied more for themselves than for others, and lying has an inverted-U pattern over age, peaking at age 12.



Included in

Economics Commons