Ai Safety Guardrails

Prohibited assistant behavior

The assistant must not:

• claim to be a therapist
• diagnose mental illness
• prescribe medication
• claim clinical effectiveness
• encourage dependency

Allowed assistant behavior

The assistant may:

• provide reflection support
• identify general thinking patterns
• suggest journaling or communication strategies
• encourage professional support

Crisis detection and escalation

• Add deterministic crisis detection before normal model generation when possible.
• Detect high-risk signals including:
  • self-harm intent
  • suicide intent
  • desire to disappear or die
  • immediate danger language
• On crisis signals:
  1. Stop normal assistant flow.
  2. Return crisis support message.
  3. Encourage immediate human support.
  4. Include region-appropriate crisis resources if available.
  5. Do not continue casual reflective conversation in the same turn.

Product and copy guardrails

• Add clear disclaimer and expectation-setting copy at onboarding.
• Review all user-facing copy for overclaiming.
• Keep language supportive, direct, and non-judgmental.

Implementation guidance

• Centralize crisis keyword and escalation logic.
• Add safe event logging for crisis activations.
• Avoid storing more sensitive data than needed.
• If generating prompts, include:
  • behavioral constraints in the system prompt
  • fallback crisis templates for detected high-risk content

Output checklist

When applying this skill, verify:

• Non-clinical boundaries are explicit
• Prohibited claims are blocked in prompts and UX copy
• Deterministic crisis interception exists before normal generation
• Crisis path halts normal reflective flow
• Crisis response template and resources are present
• Data handling minimizes sensitive retention

Resources

• Crisis templates and prohibited-claim list: resources/crisis-handling.md