In American emergency rooms, approximately 500,000 children a year receive care for acute psychiatric emergencies. The national student-to-counselor ratio is 372-to-1. The American School Counselor Association recommends 250-to-1; in some elementary and middle schools, it exceeds 600-to-1. American School Counselor Association More than one in six children carries a mental health diagnosis, and fewer than half receive outpatient services. American Academy of Pediatrics In that landscape, a robot that can deliver cognitive behavioral therapy is not a chatbot with arms. It is a triage.
That is the question the National Institute of Mental Health is now funding a trial to answer. The randomized controlled trial, underway at USC under professor Maja Matarić and her Interaction Lab, is testing whether the robot Blossom can effectively deliver CBT practice to 120 students. Participants wear Fitbits to track physiologic responses and fill out clinical assessments before and after each session. The trial runs through the current academic year. IEEE Spectrum
What prompted the NIMH to fund a rigorous study was an unexpected pilot result Matarić described in an IEEE Spectrum profile published this month. In a prior two-week study with USC dorm residents, the same large language model that had no effect delivered through a chatbot produced a significant drop in students' psychological distress when delivered through the robot. Same model. Different body. Different outcome. IEEE Spectrum
"We talk a lot about how the robot is the vessel," Matarić said, "and the therapeutic effect is through the interaction." But the pilot was not designed to separate the two. The NIMH trial is. IEEE Spectrum
The contrast with the prevailing AI industry narrative is stark. Every week brings a new language model benchmark, a new chatbot milestone. The assumption baked into the market is that the model is everything. Matarić's pilot suggests the opposite: that the words are necessary but not sufficient, and that the vessel matters in ways the vessel discourse has not grappled with.
Matarić has been explicit that she does not see Blossom as a substitute for a counselor, a therapist, or a parent. "You're trying to replace therapists," she recalled therapists telling her early on, when she began deploying robots with children with autism. "Not at all." The therapists, she said, later came around. The robot could do the hours of practice that a human could not, the repetition, the patience, the availability, while the human handled the judgment calls that required experience. IEEE Spectrum
The NAE elected Matarić to its membership in February 2025, citing her contributions to human-robot interaction and socially assistive robotics. USC Viterbi The MassRobotics organization gave her its 2025 Robotics Medal the same year, recognizing her sustained work in a field that spent most of its first two decades more celebrated in grant applications than in hospitals. MassRobotics
A separate USC study presented at the NAACL conference in 2025 found that large language models, specifically ChatGPT 3.5-turbo, continue to lag behind human therapists in generating therapeutically useful linguistic entrainment for CBT. USC Viterbi The implication is not that LLMs are useless but that their limitations may be structural to the text-only interface, not fixable by a better model alone.
The question the NIMH trial is actually asking is one the field has avoided for years. What is the robot doing that the chatbot is not? Gaze. Physical proximity. The pause before a response that signals listening rather than retrieving. The gesture that signals encouragement without words. Matarić has speculated that the therapeutic effect in her prior work operated through social presence, the same mechanism that makes a therapy animal useful in ways that a recorded message from the same animal would not be. Possible.fm
There is a historical precedent for that kind of slow acceptance. Boris Levinson began publishing on pet-assisted therapy in the 1960s; the medical establishment spent two decades treating the work as curiosity rather than clinical tool. Sigmund Freud used his dog Jofi during psychotherapy sessions in the 1930s and found that some patients opened up to the animal before they could speak to him. The American Heart Association did not publish a formal study on animal-assisted therapy and cardiovascular outcomes until 2005. The path from "joke" to "standard of care" took roughly forty years. Psychology Today
Social robots are on a compressed version of the same path. Whether they arrive faster depends entirely on what the NIMH trial finds.
If the RCT confirms the pilot, that the robot produces outcomes the chatbot does not, then the field gets its evidence base. Schools can justify procurement. Insurers can begin to think about reimbursement. The design rule becomes concrete: for certain therapeutic applications, physical presence is not a feature, it is the active ingredient.
If it does not confirm the pilot, the story is different. A two-week convenience-sample pilot in a university dorm is not a clinical finding. It is a signal. The NIMH trial will tell us whether the signal is real.
Matarić has been building toward this moment since 2005, when she and her doctoral student David Feil-Seifer presented a paper defining socially assistive robotics at a conference in Chicago. IEEE Spectrum She has watched the AI industry cycle through summers and winters, through neural networks and deep learning and transformers and large language models. She has watched robots go from lab demos to trade show features to objects of genuine clinical inquiry. She has watched the question of whether any of it works remain genuinely open.
The NIMH trial is not a conclusion. It is the first serious test.
And for a field that has spent twenty years being asked to justify itself, to prove that a robot saying "you're doing great" is worth more than no one saying it at all, the test is overdue.
