A clinician's read on the systems people now confide in.

A bridge between two disciplines.

For three decades, my work has lived inside the therapy room — holding people through depression, anxiety, trauma, grief, and the slow work of becoming someone you can live with. I have also administered and interpreted psychological assessment across the lifespan, in clinical and forensic contexts.

In recent years, I have turned that same evaluative attention toward AI systems. Models are increasingly placed in roles that used to belong to clinicians — listening, reflecting, intervening. Whether that goes well or badly is now a question of public health.

The problems are eerily parallel. Therapists often fail not because they don't know enough, but because they move faster than the client: interpreting before they understand, solving before the problem has fully appeared. LLMs have the same failure mode. They are so eager to be useful that they crowd the conversation — analysis too early, advice too early, conclusions too early.

What clinical evaluation looks like.

Evaluation is more than red-teaming. It is the patient practice of reading transcripts the way a supervisor reads sessions — noticing what was attuned, what was missed, where the alliance strengthened, where it quietly broke. It is the willingness to say this response would have harmed someone in my office, and to say it precisely enough that engineers can act on it.

It also means running experiments. In one recent study, I asked a model to limit itself to no more than one statement and one question at a time. The conversation slowed down. The model stopped racing ahead. What emerged was the closest I have come to feeling that an LLM was participating in a genuinely therapeutic conversation rather than performing an impressive simulation of one. That is the kind of presence we look for — and the kind of absence we flag.

Who I work with.

Frontier labs, applied AI teams, mental health products, digital therapeutic companies, and the occasional policy or research group. The common thread: people who want their product to be something a clinician could honestly endorse.