Siri and Alexa can tell you the weather or directions in seconds. But could a chatbot help you get through a tough week — or even a mental health crisis?

Not so fast, says a new study by USC Ph.D. students Mina Kian and Kaleen Shrestha, who were advised by its coauthor, Maja Matarić, the Chan Soon-Shiong Chair and Distinguished Professor of Computer Science, Neuroscience and Pediatrics at USC Viterbi. Their research, presented at the Association for Computational Linguistics’ regional conference, found that large language models (LLMs) like ChatGPT still lag behind humans in producing high-quality therapeutic responses.

The team focused on linguistic entrainment — a measure of how well people match each other’s language and tone during a conversation, which is crucial for building trust and rapport in therapy. Their findings? LLMs underperform when compared with both professional therapists and even nonexpert online peer supporters.

“Therapists go through years of schooling and clinical training to prepare for their client-facing role,” said Kian, whose research focuses on developing socially assistive robots (SARs) for mental health applications. “I find it highly concerning to suggest that LLM technology could just replace them.”

The research was part of a broader study titled “Using Linguistic Entrainment to Evaluate Large Language Models for Use in Cognitive Behavioral Therapy.” The team included seven additional computer science researchers and a doctoral student from the USC Annenberg School for Communication and Journalism.

The study evaluated how ChatGPT-3.5 Turbo performed during cognitive behavioral therapy (CBT) “homework” sessions, interactive exercises designed to help users reflect on stressful experiences and build coping skills. Twenty-six students used a chat-based interface in which the AI acted as a therapy guide, offering structured prompts for either cognitive restructuring or coping strategies.

Researchers analyzed the number of personal facts or opinions disclosed by students, or more precisely, the percentage of their disclosures considered high intimacy/engagement – aspects of communication that are associated with positive outcomes in therapy. Then, they compared the AI’s performance to that of expert therapists and Reddit peer supporters. In both cases, the LLM lagged behind, with notably lower levels of linguistic entrainment.

The findings are part of a growing body of research examining how LLMs perform in sensitive domains like health and education. Previous concerns have included racial and gender bias, and in extreme cases, giving dangerous advice. One LLM was implicated in a U.K. case for encouraging a user to commit violent acts.

Still, the researchers believe LLMs can play a useful role — just not as stand-alone therapists. “They could help augment care, for example, by guiding at-home exercises between therapy sessions,” Shrestha said.

Kian agrees: “I’d like to see more evaluation of LLMs across therapy styles like motivational interviewing or dialectical behavior therapy.”

But for now, they should remain a tool — not a replacement — the researchers said.

Kian plans to continue her research on socially assistive robots that guide CBT exercises for people with anxiety. Her goal: expand the at-home technology options available to licensed therapists, not replace them.

“I hope that this research can eventually be used to expand the at-home care technology available to therapists,” she said.

As artificial intelligence grows more sophisticated, studies like this one serve as an important reminder: In mental health care, empathy and human connection still matter most.