In the face of untrusted systems or untrusted people using them, a
   different kind of AI could be started with, that would first model and
   demonstrate the behaviors of itself, first, to the patient.
   Like forming a shared language before beginning a dialog.
   In such a case the patient must be aware that a malicious AI can
   convince its user of anything and craft data that makes anything appear
   true.  The user should be aware that AI forms its behaviors based on
   training data and logic, so if the AI cannot demonstrate consistent
   appropriately weighted patterns of data and logic that defend every
   tiniest part of its choices, then it is misbehaving.
   The AI can still both outcompute and deceive the user so some degree of
   trust or trust-building is needed.  People with psychosis can't trust
   all their experiences, so the process would be repeated in "symbiosis"
   with treatment.
   There is a lot of danger for taking advantage of patients: changing
   their memories permanently, etc.  We will need to find and advocate for
   each other.