Outlook: Newsletter of the Society of Behavorial Medicine

Summer 2025

Outlook main page »

Expert Eyes on the Black Box: Guiding Generative AI for Effective Behavior Change Intervention Design

Shaon Lahiri, PhD, MPH; Carol Brennan, PhD; Hanim E. Diktas, PhD, MS; Ivy Cheng, MA; John Updegraff, PhD; Erika Montanaro, PhD - Theories and Techniques of Behavior Change Interventions SIG

The era of Generative Artificial Intelligence (GenAI) to design behavior change interventions is upon us. GenAI refers to a class of machine learning models that can create new text or images based on data used to train the models.1 While still in its infancy, GenAI is being used to create personalized intervention content,2,3 gauge stakeholder views on GenAI use,4 assist in digital product creation,5 and more. Using GenAI to design behavior change interventions can potentially save time and reduce the cognitive load researchers typically expend on mundane and routine tasks. However, GenAI use also carries several risks, some of which are ethical (e.g. is this cheating?), logistical (e.g. where do I begin?), and domain-specific (e.g. is the content relevant to the population?), among others. As a theory-forward SIG, we also worry about the “Black Box” nature of GenAI models, which may potentially obfuscate the mechanisms of action in interventions

We discussed these issues with Dr. Christopher Cushing, an SBM member and Associate Professor and Director of Clinical Training in the Department of Clinical Child Psychology at Kansas University. Dr. Cushing, MPI collaborator Dr. David Fedele, and their team have been using ChatGPT to design just-in-time adaptive intervention content to increase asthma medication adherence. Dr. Cushing reflected, “We found that when we, as humans tried to create that content, it becomes repetitive, it becomes stagnant, it becomes stale, really easily.” By using ChatGPT, he found that his team was able to generate ten times as much content that would have normally taken six months to develop. Instead, with ChatGPT they accomplished the task in five minutes. “We’ve been pretty pleased with the results,” he explained to us. To accomplish this, his team first fed content into ChatGPT describing their intervention, followed by content on behavior change techniques (BCTs),6 and only then generated content in line with specific BCTs. Crucially, he then used a survey and focus groups to appraise the results with behavior change experts.

Indeed, expert appraisal is a critical component of any GenAI workflow for behavior change intervention design to achieve what Cushing calls “the right middle ground” between human-only and machine-only content creation. “GenAI can lack expertise that you really can’t identify unless you have the expertise yourself,” he clarified. Instead of devoting valuable cognitive real estate to mundane tasks, Cushing feels freed up for big picture thinking by using ChatGPT for content generation.

Cushing acknowledges that risks like data privacy require an infrastructure whose maintenance and creation are best handled by specialists. Concerns about ‘cheating’ in intervention design appear myopic to him, and need not be conflated with diminishing scientific curiosity. He further stresses that AI does not possess expertise on its own. Instead, it reflects and depends upon the expertise of the end user. Cushing invites the curious reader to “bring your skills as a scientist to GenAI and find ways to explore that curiosity with your scientific expertise.”

We thank Professor Cushing for his time.

Some ideas about a GenAI x behavior change intervention design workflow can be found here.

References:

  1. Martineau K. What is generative AI? What is generative AI? 2023. Accessed May 27, 2025. https://research.ibm.com/blog/what-is-generative-AI
  2. Sezgin E, McKay I. Behavioral health and generative AI: a perspective on future of therapies and patient care. Npj Ment Health Res. 2024;3(1):1-6. doi:10.1038/s44184-024-00067-w
  3. Harrison RM, Lapteva E, Bibin A. Behavioral nudging with generative AI for content development in SMS health care interventions: Case study. JMIR AI. 2024;3(1):e52974. doi:10.2196/52974
  4. Cudjoe T, Schoenborn N, Ashida S. Timely perspectives from diverse stakeholders on generative AI. Innov Aging. 2024;8(Supplement_1):586-587. doi:10.1093/geroni/igae098.1921
  5. Rodriguez DV, Lawrence K, Gonzalez J, et al. Leveraging generative AI tools to support the development of digital solutions in health care research: Case study. JMIR Hum Factors. 2024;11(1):e52885. doi:10.2196/52885
  6. Abraham C, Michie S. A taxonomy of behavior change techniques used in interventions. Health Psychol. 2008;27(3):379-387. doi:10.1037/0278-6133.27.3.379