AI Interprets Simulated Psychiatric Interviews, but Gaps in Clinical Reasoning Remain

Clues in psychiatric interviews lie in what patients say, as well as in their tone of voice, movements, and how their thoughts unfold. If AI can help organize this information, it could offer new tools for clinical teaching. A collaborative study by UTHealth Houston and Yale University tested this capability using simulated interview videos: the research team endorsed the overall diagnostic performance, but detailed assessments revealed gaps in reasoning that have yet to be bridged.

Published in *npj Mental Health Research*, the study used the Qwen3-Omni multimodal model alongside software developed by the team to analyze speech, audio, and behavior in the videos and generate written mental status assessments. The subjects were standardized patients portraying schizophrenia, obsessive-compulsive disorder, and bipolar disorder across different levels of symptom severity. These were controlled simulations and do not yet establish performance in real outpatient settings.

The AI and teams of psychiatric experts from the two institutions independently classified ten dimensions, including mood, appearance, speech, perception, and thought processes. According to the paper’s abstract, across 396 classifications, agreement among experts, measured by Gwet’s AC1, was 0.87, while agreement between the model and experts ranged from 0.70 to 0.72. These figures measure assessment agreement and cannot be directly interpreted as diagnostic accuracy. They also indicate that a gap remains between AI and expert judgments.

An announcement from UTHealth Houston and a university report published by Medical Xpress emphasized that the model’s overall diagnoses were consistently accurate, while also noting errors in interpreting appearance and subtle movements. The paper’s abstract further indicated that the model tended to classify observable cues, such as speech and affect, as abnormal, but was more likely to miss signs such as delusions that require understanding subjective experiences. Agreement between the model and experts also declined as symptoms became more severe.

This gives the study’s educational applications a more concrete form. The team proposed that students could compare AI assessments side by side with the judgments of experienced physicians, trace how errors occurred, and then practice deriving clinical meaning from patients’ accounts. The researchers also hope to support junior physicians or clinicians without psychiatric training in the future, with improvements across assessment capabilities as the next step.

Moving from simulated videos into clinical settings still requires validation with real patients. These results provide capabilities and error patterns that can be tested, but do not yet demonstrate that the tool can improve care outcomes. For psychiatric AI, extracting audiovisual cues is only a starting point. Whether it can reliably understand patients’ experiences and help physicians recognize blind spots in its judgments will determine how much clinical responsibility such systems can assume.