Psychotherapy has always been a very human endeavor, where the patient speaks, the therapist listens and responds, and healing occurs through words. However, with the rapid rise of conversational artificial intelligence, especially large-scale language models (LLM), that paradigm is rapidly changing.
A team of researchers at the University of Utah is working on this change, but the question isn’t: “Will robots replace therapists?” Rather, they consider more practical questions: what to automate and to what extent.
“The history of new technologies like this is almost always about collaboration and how it supports doing the work that human experts can do,” said Zach Immel, professor of educational psychology and lead author of the new study titled “A Framework for Automation in Psychotherapy.” “It might be helpful to think about a framework for understanding the different kinds of work that can be done by automation, and that’s what this paper is about.”
This study is the result of a cross-campus collaboration between researchers from the U.S. School of Engineering, School of Medicine, and School of Education.
Simply put, automation is when machines perform tasks that were previously performed by humans. In therapy, this ranges from chatbots that provide pre-written coping tips to AI systems that take and organize notes, analyze therapy sessions and provide feedback to clinicians, and even speak directly to patients.
varying degrees of automation
Co-author Vivek Srikumar uses self-driving cars as an analogy for different levels of automation.
The automotive industry has been introducing driver assistance systems to cars for many years, the most important being self-driving cars. This paper can be viewed from that perspective. The extreme version of AI in psychotherapy is an AI therapist, but there are different levels of automation and different amounts of risk involved. There may be a variety of capabilities and assistance provided by AI to therapists, clients, and organizations. ”
Mr. Vivek Srikumar, Associate Professor, Karlat School of Computing
Imel and Srikumar worked with longtime collaborator Brent Kious, associate professor of psychiatry, to create the automation framework. This framework was submitted prior to publication. Current directions in psychological science.
The team outlined four categories representing different levels of automation along a continuum.
- Category A: scripted system. The content is pre-written by a human but delivered to the patient by a chatbot that follows a decision-making tree.
- Category B: AI evaluates therapists. The AI reviews treatment sessions and provides feedback and ratings.
- Category C: AI assists therapists. The AI suggests interventions, prompts, or phrases, while a human therapist provides the care.
- Category D: AI provides direct treatment. Autonomous agents generate responses and, in some cases, interact with patients under supervision.
The team evaluated the potential usefulness and risk level of each category, which varies widely. Scripted chatbots, AI coaching tools for therapists, and fully autonomous AI therapists are fundamentally different technologies and pose different risks. However, it is often not clear which technology is being used, not only to the user but also to the healthcare system.
Weigh the risks and benefits
“By cataloging the different levels of automation, the same questions take on different flavors at different levels: questions about risk, questions about consent, questions about who consents and to what extent, the potential impact of mistakes, and questions about who and how much responsibility different parties hold,” Srikumar said. “All of these things, the questions remain the same, but the impact of these questions changes.”
The team is particularly interested in improving the way clinicians are assessed and coached to improve the level of care provided to patients.
“We are currently partnering with SafeUT, a Utah-wide text-based crisis response line, to develop a tool to help evaluate crisis counselor sessions. In doing so, crisis counselors can receive feedback to maintain important skills and also develop new skills as they learn more about crisis counseling,” Kius said.
For assessment and training, Imel says, large-scale language models can support therapists without replacing them. Current methods cannot match the scale of mental health care needs.
Automate without replacing human therapists
“Evaluating psychotherapy sessions is very labor-intensive; it is time-consuming, unreliable, and rarely used,” Immel said. “Instead of recording a session and mailing it to a specialist, the specialist can listen to the session, evaluate it, give feedback and send it back, and learn from it.” Here, a properly trained LLM can quickly grasp the core elements of treatment and provide that information to the therapist quickly, often in real time.
Researchers point out that anyone can now use ChatGPT to receive counseling similar to psychotherapy. Although LLMs are designed to sound engaging and empathetic and are trained on vast datasets, they do not necessarily use evidence-based psychotherapy techniques. Therefore, they are at great risk, as they are known to fabricate information, encode biases, and react in unpredictable ways.
“Why would you want to deploy the most risky version of a tool when you can already deploy a lighter version that makes the job easier?” Srikumar said. “For example, note-taking applications that keep notes throughout the session. These are already going to improve quality of life, quality of service for clinicians.”
The team also envisions AI playing a role in crisis hotlines one day.
“It’s a really challenging environment where you don’t know anything about the person you’re talking to. They call you and you might only have five or six talk turns to connect with them. You have a very limited space to help this person and try to keep them safe and reduce their risk,” Srikumar said. “My prediction is that future crisis counseling systems will be significantly enhanced by AI, because the scale will be too large to be satisfied without automation.”
sauce:
Reference magazines:
There it was, ZE, others. (2025). A framework for automation of psychotherapy. Current directions in psychological science. DOI: 10.1177/09637214251386047. https://journals.sagepub.com/doi/10.1177/09637214251386047
