Can AI Act as Your Therapist? Current Limitations Highlighted by Recent Study

Recent research from USC indicates that while large language models (LLMs) like ChatGPT have advanced in conversational abilities, they are not yet capable of providing meaningful therapeutic support. A study co-led by Ph.D. students Mina Kian and Kaleen Shrestha, under the supervision of Professor Maja Matarić at USC's Interaction Lab, evaluated the effectiveness of LLMs in simulated cognitive behavioral therapy (CBT) exercises.

Presented at NAACL 2025, the study revealed that these models still fall short in key aspects such as linguistic entrainment, which is crucial for building rapport and trust during therapy sessions. Entrainment refers to the responsive communication between individuals that enhances engagement and positive therapeutic outcomes. When compared to human therapists and even peer supporters on Reddit, ChatGPT displayed significantly lower levels of this linguistic adaptation.

The researchers emphasize that current LLMs should serve as supportive tools, not replacements for trained clinicians. Kian highlighted concerns over the misconception that AI could replace the nuanced expertise of human therapists, who undergo extensive schooling and clinical training. Instead, her focus is on integrating AI into mental health care to extend the reach of therapy, especially through socially assistive robots (SARs) that can support patients at home.

Participants in the study interacted with the LLM in exercises designed to help process stress through cognitive restructuring and coping strategies. While higher linguistic entrainment correlated with greater self-disclosure and engagement, the AI’s responses remained less adaptive than human interactions. The team suggests that further validation is essential to assess LLMs’ potential in diverse therapeutic modalities like motivational interviewing or Dialectical Behavior Therapy (DBT).

Kian plans to continue researching how SARs can bolster CBT homework exercises for individuals with generalized anxiety disorder. The ultimate goal is to develop supportive technologies that complement human therapy, not replace it, ensuring safe and effective mental health interventions.

Overall, current evidence underscores that AI’s role in mental health should be supportive and adjunctive, reinforcing the importance of comprehensive human care.