This randomized controlled trial evaluated the efficacy, feasibility, and technical outcomes of Amanda, a GPT-4-based chatbot, in delivering single-session relationship interventions. A total of 258 participants were randomly assigned to engage with either Amanda ( n = 130) or a writing task ( n = 128) focused on conflict reappraisal. Findings demonstrated significant improvements across 13 of 14 outcome variables—including relationship satisfaction, communication patterns, dyadic coping, problem-specific confidence, and individual well-being—over time in both conditions. Improvements emerged immediately after the intervention and were sustained or continued to improve at the two-week follow-up. However, there were no significant group differences for most outcomes, suggesting that both interventions were comparably effective. One significant group-by-time interaction emerged: participants in the chatbot condition reported lower levels of the partner-demand/self-withdraw communication pattern immediately post-intervention. The writing condition was also associated with lower overall distress about the issue. Feasibility outcomes indicated strong participant engagement with Amanda. Usability was rated highly (M = 4.19/5), as were therapeutic skills (M = 3.99/5) and working alliance (M = 4.75/6). Technical evaluation of interaction transcripts supported these findings, with high coder agreement on Amanda’s empathy, therapeutic questioning, and coherence. However, limitations were noted: Amanda occasionally produced repetitive or generic responses and did not consistently identify potential safety concerns. Overall, results suggest that Amanda provides a feasible and effective single-session relationship intervention, comparable in impact to an evidence-based writing task. This study highlights the potential for large language model-based chatbots to deliver scalable, accessible relationship support. Future research should assess Amanda’s use in multi-session interventions, explore performance in clinical populations, and enhance risk detection capabilities to ensure safe deployment in real-world settings.
Vowels et al. (Wed,) studied this question.