Julia
Elisabeth Rose.

Strategically (Im)moral: Stated Preferences in Personalized AI Advice

I study whether personalized AI recommendations based on stated moral preferences affect moral behavior. In a preregistered experiment, 1,250 subjects choose between a profit-maximizing option with no donation and a costlier moral option benefiting UNICEF. They decide themselves or delegate to a profit-maximizing or moral (donation-maximizing) algorithm. Delegation reduces moral choices by about 13 percentage points; personalized recommendations based on structured or free-text preference elicitation do not offset this decline. Structured questions elicit weaker moral preferences when subjects know their answers shape recommendations, generating more profit-maximizing advice. When subjects are not informed of this link between elicited preferences and recommendations, moral choices return approximately to the no-delegation benchmark. Personalized advice treats stated preferences as stable inputs - in moral domains, however, subjects may strategically weaken these inputs to obtain self-serving advice while maintaining a sense of following their own values. The results reveal a trade-off for AI regulation: transparency about the preference-recommendation link can distort the inputs on which personalized advice relies.

Costly Confessions
(with Fabio Galeotti)
Revise & Resubmit, Management Science

In economic interactions, lies for personal gain are common but hard to detect due to private information. In a large-scale online experiment with 1,202 subjects, individuals could confess their lies to counterparts at a personal monetary cost, which we varied across treatments. Confessions effectively helped detect lying, even when costs were high, and this effect persisted over time. However, the option of confessing neither deterred nor increased future lying, while false confessions were rare. These results support the use of confessions as an inexpensive and effective mechanism for detecting lies without affecting future behavior.