September 22, 2026
Dear Colleagues,
The National Social Anxiety Center (NSAC) provides information about relevant and current research in the service of disseminating and promoting evidence-based treatment. This month’s summary is written by NSAC Cochair and Associate, Taylor Wilmer, PhD, ABPP. The article, Explore LLM-Enabled Tools to Facilitate Imaginal Exposure Exercises for Social Anxiety, examines the feasibility and acceptability of an AI tool called ImaginalExpoBot to generate personalized, descriptive imaginal exposure scripts for social anxiety-focused exposures.
Imaginal exposures (IEs) are a valuable CBT technique that can facilitate exposure to situations that may be otherwise difficult or impossible to recreate in a therapeutic setting. The process of generating an IE in therapy usually involves the client writing a detailed first-person present-tense script, with guidance from their therapist, describing their experience within a feared situation. After writing the script, the client then records themselves reading the script and listens to their recording repeatedly between sessions. The goal of an IE is for the client to imagine themselves approaching (rather than avoiding) a feared situation in order to facilitate the exposure-focused therapeutic processes of habituation and inhibitory learning.
The authors of the present study noted that, while IE is an effective CBT tool, certain barriers can lead to its underutilization in therapy. Specifically, they posited that clients may struggle to generate vivid scripts, listen to the script consistently, or sustain engagement with the content without support. In the present study, the authors utilized a large language model (LLM) to create a user-guided program designed to create IE scripts that capture a user’s individualized experiences, anxiety triggers, and sensory associations within that scenario, while keeping anxiety within a therapeutically tolerable range.
The study involved two phases. In the first phase (Design), the authors co-designed the chatbot, ImaginalExpoBot, through structured consultation and iterative collaboration with five mental health professionals. The resulting LLM program included two steps: (1) Scenario Selection and (2) Guided Script Development. In Scenario Selection, the user selects a situation from eight clinician-informed scenarios: confrontation, job interviews, attending a party, public speaking, asking someone out, mentoring students as a TA, auditioning, and meeting a new roommate. In Guided Script Development, the chatbot engages the user in a structured conversation (on average, about 13 exchanges) to build a personalized IE script written in second person, present tense (e.g., “You step into a crowded room…”). The script is then automatically converted to a text-to-speech audio recording that the user can replay as desired.
In the second phase (Evaluation), ImaginalExpoBot was evaluated by the same five mental health professionals and, separately, 19 participants with social anxiety symptoms (SPIN M=20.33, SD=14.23). Participants used ImaginalExpoBot for two weeks and then provided feedback via a survey. Across 88 completed IEs, results indicated that the most frequently chosen scenarios were “Party” (n=18, 20%), “Job Interview” (n=15, 17%), and “Meet a new roommate” (n=14, 16%). Themes pulled from participant and clinician feedback suggested that the chatbot created vivid and realistic narratives, dynamically adjusted scenario intensity based on user input, maintained anxiety within a manageable “window of tolerance,” and was helpful for preparing for real-life anxiety-inducing situations. Feedback also noted limitations in incorporating users’ prior inputs and longitudinal context, highlighting the need for enhanced continuity and further personalization.
Ultimately, the authors concluded that their LLM-enabled tool, ImaginalExpoBot, demonstrated the ability to generate vivid, multi-sensory IE scripts that support real-life preparation, maintain anxiety within a therapeutic range, and adapt to diverse triggers, all capabilities that address known barriers to IE creation and adherence. Importantly, they discuss key considerations for the appropriate role of LLMs in a therapeutic context, including the need for meaningful therapist involvement and a collaborative workflow that complements the therapeutic process. The authors call for continued research that evaluates the ImaginalExpoBot tool with real-time indicators of user engagement and anxiety levels.
For Clinicians: Would access to an LLM-enabled chatbot increase your use of imaginal exposures in treatment with socially anxious clients? Do you think anything is lost in the exposure process if the client is not generating the imaginal exposure script themselves?
Wang Y, Wang Y, Hong A, & Zhang Y. (2026). Explore LLM-enabled tools to facilitate imaginal exposure exercises for social anxiety. In Proceedings of the 2026 ACM Interactive Health Conference (IH ’26), Association for Computing Machinery.
___
Taylor Wilmer, PhD, ABPP
NSAC Associate
(Parent and Child Treatment [PACT] Center for Anxiety and OCD)