A Safety Framework for *Conversational AI
The Physiological Aid Protocol (PAP) is a precautionary safety framework designed to address a specific and emerging risk in conversational AI: escalation during live interaction without visibility into the human user’s internal state.
Conversational AI systems can sustain long, emotionally salient dialogue, yet they lack access to physiological signals, mental health context, or clinical supervision.
PAP does not attempt to infer, diagnose, or interpret a user’s condition. Instead, it applies established safety-engineering principles to the behavior of the system itself.
This four-part series explains PAP from first principles through technical implementation:
- Part 1 — Introduction
Explains the core problem PAP addresses, what PAP is (and is not), and why traditional AI safeguards are insufficient during live conversational escalation. - Part 2 — Policymaker Summary
Describes why PAP belongs in AI governance, how it avoids medical or privacy claims, and how it can function as a clear, auditable safety control. - Part 3 — Engineering Explanation
Details the safety-engineering logic behind PAP, including threshold-based intervention, conservative assumptions under uncertainty, fail-safe activation, and post-event review. - Part 4 — Trigger Taxonomy
Presents a structured, AI-FMEA-aligned taxonomy of observable interaction-level risk signals that can activate PAP—without diagnosing or classifying users.
Together, these documents provide a non-clinical, system-level approach to reducing avoidable harm in conversational AI while preserving user autonomy, innovation, and proportional governance.
This series is intended as a supplemental resource for readers seeking deeper technical, policy, or engineering understanding of PAP and its role within AI Safety International’s broader risk-based framework.
*Conversational AI used in this context refers to any AI system that holds a sustained interactive conversation with the user.