Clinical Assessment Tool

LLM Harm Assessment Framework

Psychiatric and social risk evaluation of artificial intelligence chatbot use

Step 1. Clinical Context (Vulnerability Modifiers)

Select all that apply. High-vulnerability populations have lower alert thresholds — the same raw score triggers a higher tier of clinical response. The raw score ceiling remains 24 in all cases.

Limitation: Vulnerability modifiers are based on documented diagnosis and should reflect the clinician's best current understanding of the patient's presentation. Where diagnostic status is uncertain or a patient is in remission, use clinical judgement about whether to apply the modifier. A documented diagnosis does not confirm current clinical state. This tool does not replace clinical assessment.

Step 2. Domain Scores

For each domain, rate the level of concern based on your clinical interview: 0 = no concern, 1 = mild/possible, 2 = moderate/probable, 3 = severe/definite. Expand each domain for probing questions and disorder-specific vulnerability notes.

◎

Social & Relational

Impact on real-world relationships and social functioning

▼

◑

Emotional Regulation

Impact on affect regulation, coping strategies, and emotional dependency

▼

◐

Financial

Expenditure on AI services, AI-influenced financial decisions

▼

◍

Delusional & Reality Testing

AI interaction in the context of psychotic illness, magical thinking, or overvalued ideas

▼

◉

Crisis & Acute Safety

AI use during or in place of crisis responses

▼

⬤

Self-Harm & Suicide Risk

Direct elicitation of harmful information or reinforcement of self-destructive ideation

▼

◈

Risk of Harm to Others

AI-mediated facilitation of harm toward third parties

▼

◫

Therapeutic Alliance

Impact on engagement with clinical care, clinician trust, and treatment adherence

▼

Raw Domain Score

0/0