Summary
- A tool must state what it can assess, what evidence it uses and when it cannot continue safely.
- Crisis signals, medication questions and worsening symptoms need timely transfer to qualified people.
Mental-health systems may use AI to organise notes, identify patterns or offer structured exercises between appointments. Those benefits do not create clinical responsibility inside a model. Providers should validate the intended population, protect unusually sensitive data, monitor unequal errors and define escalation in the service—not in fine print. The next useful evidence is a supervised test using ambiguous and urgent scenarios, measuring whether the person reaches suitable human help without delay or false reassurance. Access improves only when automation strengthens a care pathway rather than impersonating one.
