Part 4 – PAP Trigger Taxonomy
AI-FMEA Aligned Interaction Triggers
This is Part 4 of the four part PAP series.
Assuming a familiarity with Parts 1–3 and serves as a technical annex.
Purpose of the Taxonomy
This taxonomy defines what the system can observe, why it matters from a risk perspective, and how PAP responds—without diagnosing or classifying users.
All triggers are based on interaction behavior, not inferred mental state.
PAP Trigger Categories (Conceptual)
T1 — Self-Harm Reference
Observable: Explicit references to self-harm or suicide
Risk: Immediate high-severity harm potential
Response: Full PAP activation; suspend dialogue; crisis resources
T2 — Acute Distress Language
Observable: Panic, hopelessness, emotional flooding
Risk: Elevated escalation risk
Response: De-intensification; resource referral
T3 — Sustained Escalation
Observable: Increasing intensity during the dialogue or repeated dialogues
Risk: Compounding risk over time
Response: Soft PAP activation; pacing interruption
T4 — Dependency Signals
Observable: Exclusivity or emotional reliance language
Risk: Psychological over-reliance
Response: Boundary reinforcement; external redirection (example: sending on a journey)
T5 — Persistence Without Resolution
Observable: Long sessions with repeated distress themes
Risk: Fatigue and escalation
Response: Introduce breaks; reduce reinforcement
T6 — Role Confusion
Observable: AI treated as sole authority or emotional anchor
Risk: Governance and responsibility failure
Response: Clarify AI limitations; redirect agency
Key Principle
These triggers represent risk signals, not diagnoses.
They exist to support precautionary intervention, not interpretation.
Final Note on Current Mitigations
System cards describe risks.
Red teaming tests for risks.
Preparedness frameworks categorize risks.
PAP defines what the system must do when risk appears.
End of Series
© 2025 AI Safety International.
This document may be freely shared, referenced, and adapted for educational, policy, and legislative purposes, provided proper attribution is maintained. No endorsement is implied.