A small pilot study presented at the 2026 American Psychological Association annual convention suggests that carefully supervised interactions with generative artificial intelligence could help people rehearse empathy, perspective-taking and relationship repair. The findings, led by Rodney Luster, PhD, a University research chair and chair of the Center for Leadership and Entrepreneurial Research at the University of Phoenix College of Doctoral Studies, point to a possible new role for AI in counselor education and guided interpersonal reflection. However, the researchers stress that the technology is not a therapist, instructor or replacement for human relationships.
Luster presented the virtual poster, “In-Session AI Reflection and Empathy: Pilot Findings and Practice Implications,” on August 8 during a virtual poster session organized by the Society for the Advancement of Psychotherapy, part of APA Division 29. The study investigated whether researcher-designed AI personas could create a consistent, transcript-rich environment in which participants practiced skills associated with empathy. These included viewing a situation from another person’s perspective, identifying emotions, responding with validation and finding language that could support relational repair.
Ten adults took part in two facilitated encounters with the AI personas. Rather than interacting with a general-purpose chatbot without supervision, participants engaged with AI systems constructed to present complex relational experiences. The facilitated format was central to the research design: participants were encouraged to slow down, consider competing viewpoints and examine how their responses affected the direction of the conversation. Researchers then reviewed the transcripts to identify patterns in communication and emotional development across the interactions.
The study used an eight-item empathy self-assessment before and after the encounters. Average scores rose from 23.10 before the sessions to 25.80 afterward, representing a mean increase of 2.70 points. Nine of the 10 participants recorded higher post-experience scores. The researchers reported a within-person effect size of dz = 0.89, a statistic that indicates a relatively large difference in this small sample. Yet the result should be interpreted cautiously because the pilot had no control group and cannot establish that AI exposure itself caused a lasting increase in empathy.
The researchers also examined how participants’ communication unfolded over time rather than focusing only on their final answers. Their analysis tracked perspective-taking, empathic concern, emotional labeling, curiosity, continuity across the conversation and meaning-making. Nine participants reached the third stage of the researchers’ relational progression framework, defined as disclosure of an underlying emotional wound. According to the reported qualitative findings, participants showed greater emotional clarity, more openness to alternative interpretations and less certainty that another person’s actions were intentionally hostile.
One potentially important observation was that progress appeared to depend on accumulated relational safety rather than a single ideal response. In other words, participants did not necessarily demonstrate empathy by producing one perfectly worded statement. Instead, their conversations appeared to develop through repeated moments of curiosity, emotional recognition and validation. This makes the transcripts particularly valuable for educators and supervisors, who could use them to examine how a person’s communication changes throughout an interaction and where an exchange begins to become more constructive or defensive.
The prospect of using AI for this purpose comes as more people turn to conversational systems while dealing with conflict, loneliness and emotional distress. Researchers and clinicians, however, have limited evidence about when AI-assisted reflection is useful, when it may be misleading and how it should be integrated into professional practice. Luster’s proposed model positions generative AI as an adjunctive process tool: a structured environment for rehearsal and reflection that remains subject to human interpretation and oversight.
That distinction is critical because AI-generated responses can be inaccurate, overly confident or insensitive to context. The presentation emphasized informed consent, privacy and confidentiality, clear professional boundaries, documentation of AI-assisted exercises and preservation of participant autonomy. Clinicians, instructors and supervisors would remain responsible for deciding whether AI-generated material is relevant, safe and appropriate for a particular person or learning situation. They would also need to review the output critically rather than treating it as an objective assessment of emotion or interpersonal behavior.
The researchers identified situations in which AI-assisted exercises may be inappropriate, including high-acuity clinical conditions, paranoia, impaired reality testing, heightened suggestibility and circumstances involving substantial privacy vulnerability. These concerns are especially significant because a simulated persona can appear responsive and emotionally fluent without possessing human understanding or clinical judgment. Future studies will need larger and more diverse samples, control conditions, longer follow-up periods and independent testing of the relational progression framework. For now, the pilot offers an intriguing but preliminary signal that supervised AI encounters might help people practice difficult conversations before applying those skills in real relationships, workplaces or clinical training. Its strongest message is not that machines can create empathy on their own, but that carefully designed technology may give human professionals another instrument for teaching, observing and refining it.
Subject of Research: People; psychological science; interpersonal skills and empathy practice
Article Title: Can Supervised AI Personas Help People Practice Empathy? Pilot Study Offers a Cautious Signal
Web References: APA 2026 program abstract and poster
References: Luster, R. “In-Session AI Reflection and Empathy: Pilot Findings and Practice Implications.” Society for the Advancement of Psychotherapy, APA Division 29 virtual poster session, 2026.
Keywords: generative artificial intelligence, empathy, perspective-taking, counselor education, psychotherapy, interpersonal communication, emotional labeling, relational repair, AI ethics, psychological science
