Assessing Chatbot Use in Psychosis
Navigating the intersection of generative AI and mental health. A synthesis of current frameworks and empirical data for clinical assessment.
The Conversational Approach
Proposed by Saba & Weeks (JAMA Psychiatry)
A motivational-interviewing-style sequence over a simple checklist. Judgmental framing suppresses disclosure.
Normalize Before Asking
Set a non-judgmental baseline.
Explore Benefits First
Patients disengage if criticized initially.
Direct Examination
Invite patients to bring actual prompts & outputs.
Ongoing Dialogue
Not a one-time screening
AI-Informed Care Framework
Morrin et al. (Lancet Psychiatry)
Management-oriented. Reframes the AI as an “epistemic ally” rather than a therapist or friend.
role: “epistemic ally”,
behavior: [
“do not claim sentience”,
“clarify limitations”,
“refer to clinical care”
],
// Requires co-design with service users
Domains Supported by Empirical Data
Highest-yield screening questions based on the Buck & Maheux survey.
Ever-use alone does NOT differentiate risk groups.
Intensity
Several times daily, >30 min/day, ≥6 conv/day
Role Ascription
Viewing AI as a companion, friend, therapist, or romantic partner.
Clinical Content
Delusion-related interactions (13.3%-30.7% in elevated risk).
Functional Typology
Flathers et al. (Lancet Digital Health)
Classifying system role to guide clinical/tech responses, moving away from a unitary label.
Adaptable Existing Frameworks
Psychiatric Inquiry into Digital Media
Moreno et al. (Psychiatric Services)
Emphasizes quality over quantity. The closest general-purpose model.
Problematic Internet Use Scales
ISAAQ, PIUQ-6, Young's IAT
Captures compulsivity, but misses AI specific risks (anthropomorphism/delusion). Partial proxy only.
Technology Use Survey
Dwyer et al.
Assesses digital literacy, not risk. Good for checking if patient understands what an LLM is.
The Defensible Current Standard
Currently, none of these tools are validated against psychosis outcomes. Only 16% of LLM chatbot studies undergo clinical efficacy testing (Hua et al.). The practical standard is to incorporate a brief structured inquiry into routine assessment and document it as a maintenance factor where positive:
Keep reading
All understanding psychosis pages →- Understanding psychosisAI-associated psychosisHow conversational AI can amplify aberrant belief formation.For everyone
- Understanding psychosisConversational AI & psychosis riskRisk factors when chatbots become a primary reality check.For everyone
- Understanding psychosisAI psychosis: surveillance reportTracking reported cases and emerging patterns of AI-associated psychosis.For clinicians
- Understanding psychosisWhat is psychosis?Psychosis, thought disorder, and schizophrenia explained from the ground up.For families
- Understanding psychosisLearning about psychosis & thought disordersA guided tour of the psychosis spectrum and the words clinicians use.For families