Safety, efficacy and acceptability of human-GenAI single-session exposure-based intervention for academic anxiety: randomized controlled trials

Generative AI (GenAI) may expand access to exposure therapy, a gold-standard anxiety treatment, but its safety, efficacy, and acceptability remain untested. We developed a hybrid human–GenAI single-session exposure program targeting academic anxiety and evaluated it across two randomized controlled trials (N = 425; Study 1: ChiCTR2500101784, preregistered April 29, 2025; Study 2: OSF preregistered August 13, 2025, https://doi.org/10.17605/OSF.IO/DW9B2; ChiCTR2600123876, retrospectively registered April 30, 2026) using waitlist and active control. The program was perceived as safe and acceptable by participants.

Compared with both controls, the intervention generally produced greater improvements across primary outcomes, including immediate state anxiety and two-week follow-up academic anxiety and procrastination (d = 0.25–0.35 vs. active control; 0.43–0.58 vs. waitlist). Effects on secondary outcomes, including positive and negative affect, generalized anxiety, and depressive symptoms, were limited. These findings provide initial evidence that human–GenAI collaboration can safely deliver exposure-based interventions, potentially expanding access to evidence-based care.

This project was funded by the National Natural Science Foundation of China (Project Numbers 32200907, U2336214), Tsinghua Initiative Scientific Research Program (2022THZWJC12), and the Beijing Municipal Education Commission (BJSZ2025ZD02). The funders played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript. These authors contributed equally: Yanling Yue, Yanwen Yu.

Department of Psychological and Cognitive Sciences, Tsinghua University, Beijing, China Yanling Yue, Yanwen Yu, Chuncheng Huang & Wenting Mu Department of Computer Science and Technology, Tsinghua University, Beijing, China Department of Psychology, University of Illinois at Urbana-Champaign, Champaign, IL, USA School of Kinesiology and Health, Capital University of Physical Education and Sports, Beijing, China Correspondence to Nisha Yao or Wenting Mu. The authors declare a patent application related to this work: “Method and Device for Dialog Control in Psychological Interventions Based on Large Language Models” (application number: 202610587633.X). Tsinghua University and Capital University of Physical Education and Sports are the patent applicants.

At the time of submission, the application remains pending and has not yet been published. The inventors are W.M., Y.Yu., A.W., N.Y. and J.L., of whom W.M., Y.Yu. and N.Y. are authors of this manuscript. The patent covers aspects of the chatbot-guided GenAI-augmented exposure component described in this manuscript.

The other authors declare no competing financial or non-financial interests. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. During the preparation of this work, the authors used generative AI technology, specifically OpenAI’s ChatGPT-5, to enhance the readability and linguistic presentation of the manuscript, as well as to proofread and translate excerpts from interviews from Chinese to English.