Financial_choices_and_advice_v1

This is a second update to the first registration. For some reason, if I click "Update" on the first registration, it does not allow me to upload files. Data collection has not yet begun. This update reflects changes to the lab setup (single room with alternating batches instead of two simultaneous rooms), batch scheduling procedure, compensation (base payment updated to £15), addition of a 17th bias (T — Default Effect, bringing total trials to 27), minor prompt wording refinements, and documentation of standardised informational aids (Important Box and Short Glossary) shown to all participants. Description: This lab experiment tests whether access to a GPT-based AI advisor affects behavioral biases in financial decisions. Participants (N=120) are run in a single room in batches of 15, alternating between GPT (AI chat available) and NO_GPT (no AI) conditions. Each participant completes 27 trials measuring 17 cognitive biases (e.g., loss aversion, framing effect, anchoring, sunk cost fallacy, base rate neglect). Research Question: Does GPT assistance reduce or amplify biases? Hypothesis: Non-directional — we test whether behavior differs between groups without predicting direction. Primary Analysis: Compare GPT vs NO_GPT responses for each bias using chi-square, t-tests, ANOVA, and logistic regression as appropriate. Secondary Analysis: Per-protocol analysis (GPT participants who actually used the AI), FDR-adjusted p-values for multiple comparisons, and within-person contrasts for mixed-design biases. Exploratory Analysis: GPT usage patterns, demographic moderators of GPT effects, and composite bias scores. Ethics: LBS REC1085