Abstract
Cognitive Bias Modification-Interpretation (CBM-I) is an experimental paradigm that has been adapted into a digital intervention for various mental health concerns, including depression. Developing training materials for CBM-I can be resource-intensive. Recent findings suggest AI-generated CBM-I materials may approximate the therapeutic content of human-generated items. Here, we investigated whether CBM-I using AI-generated scenarios reduces negative interpretation bias and compared its effects with human-generated materials and a control condition. Participants (N = 165) endorsing mild range or higher depressive symptoms on a screening measure were randomly assigned to one of three groups: CBM-I running AI-generated items (n = 54), human-generated items (n = 55), or a text-reading control (n = 56). Each participant completed a single-session training. Interpretation bias and state affect was measured at pre-and post-training. Results showed that both CBM-I groups significantly reduced negative interpretation bias relative to control; non-inferiority analysis revealed that AI-generated materials were not worse than human-generated materials. Results were not triangulated across the two interpretation bias measures, however. The findings provide proof-of-principle evidence that generative AI could create CBM-I training materials without compromising fidelity. Given notable limitations, however , future studiesusing scheduled delivery and long-term follow -ups are needed to establish the clinical significance of AI materials.