Social desirability bias (SDB) is a pervasive threat to the validity of survey and experimental data. Respondents might often misreport sensitive attitudes and behaviors to appear more socially acceptable. We begin by synthesizing empirical evidence on the prevalence and magnitude of SDB across various domains, focusing on studies with individual-level benchmarks. We then critically assess commonly used strategies to mitigate SDB, highlighting how they can sometimes fail by creating confusion or inadvertently increasing perceived sensitivity. To help researchers navigate these challenges, we offer practical guidance on selecting the most suitable tools for different research contexts. Finally, we examine how SDB can distort treatment effects in experiments and discuss mitigation strategies.

 We test the effectiveness of different AI-delivered conversation protocols to increase people's motivation for change. In a large-scale experiment with 2,719 social media users, we randomly assign participants to a control conversation or one of three treatment arms: two Motivational Interviewing protocols promoting self-persuasion (change focus or decisional balance) and a direct persuasion protocol providing unsolicited advice and information. All conversations are led by an AI interviewer, enabling standardized delivery of each protocol at scale. Our results show that all three interventions significantly increase motivation for change and the perceived costs of social media use, with change-focused self-persuasion yielding the largest effects. These effects persist and translate into self-reported reductions in social media use more than two weeks after the intervention. Our findings illustrate how AI-led conversations can serve as a scalable platform both for delivering behavioral interventions and for testing what makes them effective by systematically varying how conversations are conducted.

 InA large share of those affected by mental illnesses go untreated. Psychotherapy, a first-line treatment, works through a collaboration between patient and therapist, and that relationship is why many stay away: therapy is expensive, and to be helped, one must disclose to another person and risk being judged. In a randomized controlled trial with 400 US adults with moderate-to-severe anxiety and depression, none in therapy and most of them lonely, we replace the human side of the collaboration with a voice-based AI delivering CBT. For this population, the treatment appears to survive the removal of the therapist: participants take up the AI-delivered care and their demand shifts away from the human provider, not away from care or social support. Assignment lowers anxiety and depression by roughly 0.8 standard deviations at four weeks, within the range reported for human-delivered CBT, moves the mechanisms CBT targets, and forms a working alliance matching published norms for human therapists. Demand for human therapy falls on a stated measure and, by 18 percentage points, in a behavioral choice, while demand for medication and reliance on friends and family are unchanged; among participants who prefer the AI, the stated appeal is the absence of another person’s judgment. Treated participants also report more social contact, not less, and lower loneliness, with the contact gains concentrated among the loneliest. By removing the person, automation may expand access to relational services, lowering both their price and the social cost of seeking them.

with Lukas Wolf, Valerie Forman-Hoffman, Patricia Areán, and Bennet Feld

Elevated anxiety symptoms are common, yet many people lack adequate support because professional care is costly, scarce, and stigmatized. AI-based systems may help overcome these barriers by providing scalable, on-demand support. In this study, we evaluated Sonia, a voice-based generative AI companion teaching techniques from cognitive behavioral therapy. Four hundred US adults with elevated anxiety symptoms were randomly assigned 1:1 to Sonia or passive web-based psychoeducation based on World Health Organization materials. The pre-registered primary outcome was the intention-to-treat standardized Generalized Anxiety Disorder-7 (GAD-7) score at 2 weeks. Sonia participants had GAD-7 scores 3.52 points lower than controls (95% CI [-4.42, -2.62]). Five participants raised safety concerns; 3 met pre-specified risk thresholds and were directed to crisis resources. These findings suggest that voice-based AI companions such as Sonia may provide an accessible complement to existing care.

 Private information is often viewed as a source of bargaining power. We argue that its value depends on how bargaining translates hidden values into prices. We study this link in a model-guided, continuous-time laboratory experiment with 384 subjects and 5,393 bilateral negotiations. The experiment varies information structure and transaction costs while preserving an unstructured bargaining protocol. Bargaining reveals private information only partially. Rather than separating through delay, subjects rapidly exchange offers and narrow the gap between opening positions through successive revisions. This partial revelation generates price compression: deal prices respond to private values, but less than full revelation predicts. Consequently, private information has different distributional effects across types. High-value informed buyers earn information rents, whereas intermediate types are harmed. In our setting, private information becomes bargaining power only when bargaining translates information into prices with sufficient precision.