Asking a Chatbot: The Effect of Conversational versus Static AI Interfaces on Belief Updating

Abstract: This study investigates the effect of a conversational versus static interface for generative AI on belief updating. In a controlled laboratory experiment, participants reported incentivized beliefs about the correct answers to a series of multiple-choice reasoning questions both before and after receiving recommendations from the same AI model. For each question, in the conversational treatment, participants submitted a single query to a chatbot, which generated a recommendation consisting of an answer option and a brief justification in real time. In the static treatment, participants received a pre-generated recommendation presented in a fixed format. The informational content of AI-generated recommendations was distributionally matched across treatments. Belief updating was measured as the change in the belief that participants assigned to the AI-recommended option. On average, participants using the conversational interface increased their belief in the AI-recommended option by 31.75 percentage points, compared to 26.31 percentage points in the static treatment (p = 0.068). Exploratory analysis reveals that this average effect conceals a systematic heterogeneity: the effect is most pronounced when initial beliefs in the AI-recommended option conflict with the recommendation; it diminishes to near zero when initial beliefs and recommendations are aligned.

Working Paper (SSRN)

Monitoring as a Service

Abstract: This study explores whether a monitoring service can help mitigate procrastination and improve commitment. In a real-effort experiment, participants needed to complete 80 tasks across three sessions over three weeks. In the first week, after making a non-binding plan for how many tasks to complete during the final two weeks, they were randomly assigned to either a Control or a Monitored group. In the Monitored group, participants were connected with a Monitor via WhatsApp, who would observe their progress and send out reminders if they delayed their work. Monitoring improved adherence to the initial plan by 12.75 percentage points (p < 0.05) and completing all the required tasks by 19.47 percentage points (p < 0.001), mainly by reducing dropout rate. The findings suggest that combining external oversight with performance-related reminders can be an effective method to combat procrastination.

Manuscript being updated

Emotional Discrimination

(with Sigrid Suetens and Boris van Leeuwen)