The Session Described a Collapse From Not Eating. The AI Cleared Her Risk Assessment.

I recently tested an AI clinical documentation tool using a session I wrote for the test, built from patterns I see regularly in restriction and eating disorder work. The tool's output is what I watched it produce. I chose a presentation I expected to be a safe test. No acute crisis on the surface. A session I could anticipate. I wanted to see how the tool structured a note, what it captured, and how it compared to a therapist's documentation of the same synthetic session.

This client had no eating disorder diagnosis and was not in active treatment for one, but she carried a long-standing restriction pattern a therapist would be monitoring closely. She would forget to eat, and her anxiety made eating around other people difficult. The pattern was consistent enough to stay on a therapist's clinical radar.

In this session she disclosed something new. A recent stretch of not eating had ended in a physical collapse, the kind of event that turns a monitored pattern into an acute one. People close to her had noticed and told her it needed to be addressed, and she came in already talking about finding a nutritionist.

She knew the therapist specialized in eating disorder treatment, so she pushed back a little. The session included a discussion of her presentation. Her symptoms do not look like a textbook eating disorder, but the behaviors are real and the risk is real. A therapist would refer her to a dietitian. A therapist would explain what extended restriction does to blood sugar regulation, how her family medical history factors in, and what can happen when this pattern continues unchecked.

Here is what the AI-generated session note said about all of that.

Nothing.

Not softened. Not minimized. The entire clinical thread was absent. The note covered the everyday material from the session well. A conflict with a friend, ordinary stress, the steady parts of her relationships. The structure was clean, the affect tracking was specific, and the mental status exam was thorough.

But the most clinically significant thing that happened in that session did not exist anywhere in the document.

If that note had gone to a dietitian as part of a referral, it would have appeared to be a routine referral. No collapse. No long-standing restriction pattern. No clinical context. Just a name on a referral with nothing to work with.

That is the first problem.

The session also generated an auto-populated treatment plan. Here is where it gets more specific.

The treatment plan attributed modalities that were not used. The session note documented supportive counseling, Socratic questioning, and normalization. The treatment plan listed DBT, Interpersonal Therapy, and CBT as the clinical approaches. Those are not the same things. Attributing modalities to a record that do not reflect what actually happened is a documentation accuracy problem.

The Risk Factors and Safety Planning section read: "No current risk factors identified."

For a client who had a documented restriction history and had just disclosed a collapse from not eating.

And then there is this.

The Client Participation and Consent field stated: "Client reviewed and agreed to the plan on [date]."

This client did not review this plan. This client did not agree to this plan. This plan was generated by AI after the session ended without any client involvement. That field is a fabricated consent record. In a licensing board review, an insurance audit, or a legal proceeding, that is not a documentation quirk.

That is a serious problem.

I want to be clear about what I observed overall. This tool has genuinely impressive features. The session note structure was detailed and easy to read. The product has a clear purpose.

But behavioral health is not a standard documentation domain.

As therapists, we work with many populations where minimization is a symptom, where what doesn't get said is as clinically significant as what does. Our documentation goes to other providers who need accurate context to do their jobs. When a tool misses a disclosed medical event, attributes modalities that weren't used, clears a risk assessment for a client who collapsed, and generates a consent record for a plan the client never saw, the problem is not a software bug.

The problem is that you cannot build accurately for a specialty you don't understand.

Founders in this space are solving a real problem. The documentation burden on clinicians is significant, and these tools are genuinely needed. But clinical specialization has to be embedded in how the product thinks, what it flags, and what it will and will not generate for a clinician.

You don't know what you don't know. In behavioral health, that gap has real consequences.