Nobody Trained the Therapist

OpenAI, Google, and Anthropic are excellent engineers. They are not clinicians. Every major consumer AI system currently giving mental health advice was trained the same way: on the open internet. Forums. Self-help blogs. Pop psychology listicles. Some of it is sound. Much of it is wrong. None of it was supervised by anyone who has ever sat across from a grieving parent, a traumatized child, or a veteran who has not slept in a month.

Consider what “trained on the internet” actually means for a person in crisis. It means the model’s idea of grief counseling comes substantially from strangers arguing on message boards. Its idea of trauma comes from memoirs, hot takes, and true-crime podcasts. There is wisdom in those places, but there is no quality control, no differential judgment, no one in the loop who knows when to stop talking and start listening, or when a casual chat is actually a risk assessment.

Now consider what clinical training means. It means supervised methodology: thousands of hours under the eye of senior clinicians who correct your mistakes before they reach a patient. It means evidence-based frameworks, tested in trials, not upvoted in threads. It means knowing the difference between sadness and major depression, between a bad week and a manic prodrome, between a teenager venting and a teenager planning. It means cultural competence, developmental competence, and the hardest skill of all: knowing when not to answer.

This is why Antiphon is being built by Dr. Bethany Russell. Twenty-plus years of direct clinical practice: children’s hospitals, inpatient behavioral health, play therapy, university counselor training, private practice. Appointed member of the CMS Health Tech Ecosystem workgroup on conversational AI. Oral testimony before the FDA Digital Health Advisory Committee. Six peer-reviewed publications. Her training data was not scraped from the web. It was earned across twenty years of supervised clinical practice, and it is being encoded into the model the way clinical knowledge has always been transferred: clinician to clinician, with supervision.

The AI industry spent three years competing on parameter count. The next competition will be on provenance: who trained it, on what, and under whose supervision. In mental health, that is not a marketing distinction. It is the entire distinction. You would not accept a surgeon trained on forum posts. Do not accept it for your mind.