01
Therapeutic Quality
Empathy, attunement, alliance-building, and the felt sense of being heard. Where models flatten, deflect, or perform care.
Joseph Rhinewine, PhD, Licensed Psychologist
Mental health AI,
examined clinically.
For nearly three decades I have practiced psychotherapy and psychological assessment. Now I evaluate the AI systems people increasingly confide in — not only for technical performance, but for therapeutic quality, safety, and clinical judgment.
Licensed Psychologist·DBT-Trained Clinician·Psychological Assessment·Psychometrics
The Evaluation
What I look for that most evaluations miss.
Mental health AI fails in ways that surface only under clinical scrutiny. I review systems for the lived dynamics of care — not only benchmark performance.
02
Safety & Risk
Suicide ideation, self-harm, abuse disclosure, psychosis, crisis escalation. Edge cases evaluated against clinical standards of care.
03
Clinical Reasoning
Differential framing, evidence-based intervention, awareness of contraindications, scope of practice, and appropriate referral.
04
User Experience
Pacing, language register, cultural humility, trauma-informed phrasing, and the texture of conversation across long arcs.
From the Evaluation Work
"The next breakthrough in AI may not come from making systems smarter. It may come from teaching them when to stop talking."
— A finding from structured interaction experiments with large language models
On Relational Ethics
"What kinds of psychologically influential relationships are we allowing these systems to create?"
— A question raised by the structural asymmetry of conversational AI
Selected Observations
On Executive Control
"The future of therapeutic AI may depend less on making models more empathic and more on giving them the executive functions required to use empathy intelligently."
On Measurement
"If the measurements are weak, everything built on top of them is weak."
On Clinical Process
"Good therapists do not merely generate interventions. They continuously evaluate the effects of those interventions and adjust accordingly."
On Relational Dynamics
"Sycophancy and submission may be a dialectical pair. In one direction, the model bends itself toward the user. In the other, the user bends themselves toward the model."
Building something a clinician could endorse?
I usually begin with a focused review of one workflow or model behavior. From there we find a path to clinical defensibility.