Feedback beats practice
- Stanford and Georgia Tech researchers reported that CARE, an artificial intelligence counseling trainer, improved novice counselors’ empathy only when feedback accompanied chatbot practice. - In a randomized study of 94 novices, the feedback group improved reflections and questions, while practice alone worsened empathy over time. - The result fits a broader Stanford Human-Centered Artificial Intelligence push toward pairing simulated practice with structured coaching, not standalone chatbots. (arxiv.org) (hai.stanford.edu)
Learning counseling is partly about hearing what to say next, and partly about hearing what you missed. Stanford and Georgia Tech researchers say the second part is where artificial intelligence made the difference. (arxiv.org) In their CARE system, novices practiced with an artificial intelligence “patient” that played out therapy-style conversations. Some users got only that role-play, while others also got an artificial intelligence mentor that critiqued their responses and suggested better ones. (arxiv.org) (hai.stanford.edu) The randomized study enrolled 94 novice counselors. The practice-and-feedback group improved in its use of reflections and questions, with effect sizes of 0.32 to 0.39 and p-values below 0.05. (arxiv.org) The practice-only group did not improve on those measures. On empathy, it got worse over time, with an effect size of 0.52 and p=0.001, and it also underperformed the feedback group by 0.72 standard deviations. (arxiv.org) That split showed up in how participants described what they learned. Users who got feedback moved toward a client-centered style built on listening and validation, while practice-only users stayed more focused on solving the client’s problem. (arxiv.org) The distinction matters because counseling skills are not just scripts. Reflections, open questions, and validation are the tools therapists use to keep the conversation on the client’s experience instead of jumping straight to advice. (arxiv.org) (hai.stanford.edu) The researchers place CARE inside a broader “AI Partner, AI Mentor” model. The partner gives people a safe practice environment; the mentor turns that practice into instruction by pointing out mistakes, naming goals, and offering revised responses. (arxiv.org) (hai.stanford.edu) Stanford’s earlier peer-counselor work used the same logic. Researchers worked with three Stanford psychotherapists to define useful feedback, then built a system that explains the counseling goal, suggests improvements, and proposes a stronger reply. (hai.stanford.edu) (aclanthology.org) The paper starts from a capacity problem in mental health care. It notes that 22.8% of U.S. adults, about 58.7 million people, experienced mental illness in 2023 while training skilled helpers remains slow and labor-intensive. (arxiv.org) The closing lesson is narrower than “artificial intelligence can teach empathy.” In this study, practice built confidence, but feedback changed behavior. (arxiv.org)
