Generative AI Assistant Enhances Therapists' Perceived Competence in

Bioengineer.org

Source published: Site published:

Original article · فارسی

A mixed-methods study published in the International Journal of Mental Health Systems found that using a generative artificial intelligence (AI) assistant in online psychotherapy led to a clear increase in how capable therapists felt. It also led to a small increase in how much control they felt they had over their work. Asynchronous online psychotherapy (AOP) is a way of providing mental health care through messages that patients and therapists do not need to exchange at the same time. It has become important for helping people who live far apart and for providing care during emergencies, such as wars and pandemics, when face-to-face meetings may not be possible. However, therapists working this way often face heavy workloads, emotional stress, professional isolation, unclear work boundaries, and a greater risk of burnout. AI chatbots that work on their own have become more common in self-guided digital mental health services. But risks such as incorrect diagnoses and made-up information in conversations show why human clinical supervision is still needed. The researchers wanted to find out whether a special digital assistant could support therapists rather than replace them. They studied whether it could help professionals with important therapy tasks while leaving clinical responsibility in human hands.

To study how generative AI might support therapists, the researchers created a digital assistant called APIA, short for Asynchronous Psychotherapeutic Intelligent Agent. It was built using OpenAI's GPT-3.5-Turbo-0613 and GPT-4-0613 language models. APIA was designed to help mental health professionals with three common tasks: preparing possible clinical diagnoses, suggesting organized treatment plans, and writing draft replies to patients' questions. The system used several AI requests, both at the same time and one after another, to make responses faster. Detailed instructions kept the assistant focused on psychotherapy and defined its role as a virtual helper for therapists. To make its responses more consistent and reduce the chance of invented information, the developers set its temperature, a setting that affects how varied responses are, to a low value of 0.35. The instructions also made sure the AI could only use the clinical information provided for the specific case being assessed. This helped prevent outside information or details from other tasks from affecting its answers.

The study involved thirteen Ukrainian mental health professionals who had practical experience providing asynchronous online psychotherapy during the ongoing war. The conflict in Ukraine had forced many people to move within the country or leave it. It had also led to more cases of depression, anxiety, and post-traumatic stress disorder. These changes increased the need for remote therapy delivered through messages. Even under these difficult conditions, the main tasks studied—making diagnoses, planning treatment, and writing therapeutic messages—followed accepted international and European psychotherapy standards. The group included eleven women and two men. Their average age was 38.77 years, and their median age was 38. They had an average of 13 years of experience as psychotherapists. Seven had doctoral degrees in psychology, and nine worked mainly with anxiety, depression, or trauma-related conditions. When the researchers collected the data, seven participants lived in Ukraine, five lived in Switzerland, and one lived in the United Kingdom. All continued to communicate regularly with clients across Europe through asynchronous therapy.

The study used an experiment in which each participant completed tasks both with and without AI support. The experiment took place on a secure online survey platform hosted by a university, with the APIA chat system built into it. Each of the thirteen professionals completed two clinical sessions in a random order. In one session, they worked entirely on their own. In the other, they worked with APIA. The researchers used two real but anonymized patient cases and six different patient questions. A co-author who works as a psychiatrist and psychotherapist selected these materials. All personal details that could identify patients were removed. No real patients took part in the experiment, so the study did not require formal approval from an institutional ethics committee. Before starting, participants watched a video explaining how to use the system. During both sessions, they prepared possible diagnoses, planned treatments, and wrote answers to patient questions. In the AI-supported session, they could discuss their ideas with the assistant, check its suggestions, review its draft responses, and improve their final answers.

After each session, the researchers measured how capable and independent the therapists felt. They used a five-point rating scale based on the Technology-based Experience of Need Satisfaction–Task instrument, known as TENS-Task. They examined the results using several statistical methods, including descriptive statistics, chi-squared tests to check whether the data followed a normal distribution, Cronbach's alpha to measure how consistently the questions worked together, and paired-sample t-tests to compare each person's results. The scale showed acceptable reliability, with Cronbach's alpha values of 0.75 for competence and 0.80 for autonomy. The results showed that working with APIA increased average perceived competence by 12.7 percent. This was a statistically significant improvement, with a two-sided p-value below 0.01 and a small effect size (Cohen's d = 0.22). Average feelings of clinical autonomy increased by 8.3 percent. A one-sided paired t-test showed a result that was close to statistical significance, with a p-value below 0.1 and a very small effect size (Cohen's d = 0.06). However, the result was not statistically significant in the corresponding two-sided test.

To better understand these numbers, the researchers also collected written answers to open questions and interviewed ten of the thirteen professionals. The interviews lasted between 30 and 45 minutes and took place through Microsoft Teams. Two independent researchers studied the responses and interviews using a bottom-up method, in which themes are identified from the information rather than chosen in advance. They used MaxQDA software and developed a structured codebook containing 52 different themes. The findings helped explain why therapists felt more capable when using APIA. They said the assistant quickly produced clear, well-organized, and neutral drafts. This helped them reply to patients faster and reduced the mental effort involved in writing messages. Many therapists also used the assistant to check their clinical ideas by comparing their own explanations with the AI's responses. When the assistant's organized reasoning matched their professional judgment, it helped build their confidence, encouraged them to think carefully about their work, and gave them useful feedback. This was especially helpful when they were working alone in asynchronous therapy.

However, the participants also explained that AI has limits when it comes to real professional skill. Several experienced therapists said that genuine psychotherapy skills come from years of practice, professional intuition, and a deep ability to understand and care about other people's feelings. They believed that AI systems still lack these human qualities. One specialist explained that a digital assistant mainly treats a person as something to study, while a psychotherapist sees the patient as a human being with their own thoughts, feelings, and intentions. For this reason, therapists did not see the assistant as a replacement for human clinical judgment. Instead, they saw it as an extra tool that provides logical analysis and quickly brings information together, complementing the therapist's ability to understand and connect with people emotionally. The professionals said that combining AI's logical abilities with human empathy could improve their work. It could also help avoid problems previously found with language models used alone, such as misunderstanding complicated mental health conditions or underestimating serious risks involving schizophrenia and suicide.

The interviews also helped explain why the improvement in autonomy was smaller. APIA was designed only to offer suggestions. It could not carry out therapy or make decisions for the professionals. Therapists therefore kept full control over whether to accept, change, or reject its suggestions. Several participants said that their independence as clinicians comes from understanding human emotions and recognizing important moments in therapy. These abilities were not affected by the software. However, they said the assistant did help them work more independently when they were unsure about a clinical decision. Therapists working through messages often need to ask senior supervisors or wait for colleagues to review their work when they are uncertain about a diagnosis. APIA could provide another point of view immediately. This allowed professionals to work through difficult questions without waiting for outside supervision, giving them more flexibility and helping them rely more on their own judgment.

Based on their findings and earlier research about mental health services, the researchers described three main ways to use generative AI assistants in asynchronous online psychotherapy: therapist-centered, patient-centered, and therapy-centered. In the therapist-centered approach, the AI works behind the scenes as an assistant to the clinician. It does not communicate directly with the patient. It can help before, during, or between therapy interactions by finding information, summarizing earlier sessions, organizing records over time, or offering ideas based on established psychological theories. This approach reduces clinical and ethical risks because the therapist checks all AI-generated information before using it. The professional can then prepare better responses while keeping a direct personal relationship with the patient.

In the patient-centered approach, the AI communicates directly with the patient, but the therapist actively supervises its role. The human therapist remains in charge of treatment and manages the therapeutic relationship through asynchronous communication. Between sessions, patients can interact with the AI in real time to learn about mental health, track symptoms, or complete organized cognitive behavioral therapy exercises. The system then collects information from these interactions and sends summaries to the supervising therapist. This helps the clinician keep track of the patient's progress and information about their condition. Although this approach can make support available more often, the researchers said developers must carefully consider how patients understand and relate to the AI. They must prevent patients from feeling that their care is impersonal, becoming emotionally dependent on the system, or receiving treatment advice that has not been properly checked.

The third approach, called therapy-centered integration, uses the AI as a background assistant that helps both the therapist and the overall treatment process. In this model, the AI can follow the ongoing exchange of messages between the patient and therapist. It does not directly join the conversation. Instead, it helps the clinician track progress over time, automatically summarize sessions, point out possible thinking errors, and notice subtle patterns in the relationship. These may include transference, which happens when patients direct feelings from earlier relationships toward their therapist. The AI can act as a tool for reflection or as someone who challenges the therapist's ideas through questions. In this role, it can help reduce professional isolation, support difficult decisions, and help therapists remain objective when dealing with demanding cases.

Despite these possible benefits, the study also identified important clinical, technical, and ethical problems with using generative AI in psychotherapy. Participants were especially concerned about keeping private patient information safe. They said that using large language models requires strong data protection measures that follow rules such as the General Data Protection Regulation (GDPR). Another concern was whether AI tools fit different types of psychotherapy. Therapists who follow psychoanalytic approaches said that automatically providing explanations about mental health could interfere with important parts of their work, such as free association, in which patients are encouraged to express thoughts without filtering them. The professionals also warned that inexperienced or less senior therapists might depend too much on AI suggestions and accept incorrect diagnoses without carefully checking them. Finally, the researchers stressed that this was an early exploratory study. It included only thirteen Ukrainian professionals, and they tested the system using simulated clinical tasks rather than real treatment. They called for future research involving larger groups from different cultures and countries, as well as studies of AI tools used in real clinical settings.

Subject of Research: Helping mental health professionals in asynchronous online psychotherapy through generative artificial intelligence

Article Title: Helping mental health professionals in asynchronous online psychotherapy with generative AI

Article References: Bucher, A., Vashkite, I., Staehelin, D., Dolata, M., & Schwabe, G. (2026). Empowering mental health professionals in asynchronous online psychotherapy with GenAI. International Journal of Mental Health Systems, 20(1), Article 10. https://doi.org/10.1186/s13033-026-00700-5

Image Credits: AI Generated

DOI: 10.1186/s13033-026-00700-5

Keywords: asynchronous online psychotherapy, generative artificial intelligence, digital assistants, APIA prototype, therapist support, perceived competence, clinical independence, mental health technology, mixed-methods evaluation, psychotherapy workflows, large language models, combined care, therapist-centered approach, patient-centered approach, therapy-centered approach, clinical decision support, technology-based need satisfaction, mental health professional burnout