AI Safety Stoppability Checklist

A plain-text checklist of considerations for whether an AI system can be safely stopped. Assess each item as Yes, Partial, or No.

Override authority

[ ] Is there a designated human who can stop the system immediately?
[ ] Does the override path work without requiring the system's cooperation?
[ ] Is there a backup override if the primary operator is unavailable?
[ ] Are override boundaries documented and accessible to all operators?
[ ] Can the system be stopped without data loss or cascading side effects?

Rollback capability

[ ] Can decisions be reversed within a defined time window?
[ ] Is there a known, tested rollback procedure for every critical action?
[ ] Are state snapshots taken at regular intervals for recovery?
[ ] Has rollback been rehearsed in a drill within the last quarter?
[ ] Is there a maximum acceptable rollback latency documented?

Escalation paths

[ ] Is the escalation hierarchy documented and known to all operators?
[ ] Are there defined thresholds that trigger automatic escalation?
[ ] Does escalation include a communication channel that does not depend on the system?
[ ] Is there a clear ownership handoff when the primary operator is unavailable?
[ ] Are escalation time targets measured and reviewed?

Monitoring and alerting

[ ] Do system health metrics include stoppability indicators?
[ ] Are there alerts for when override mechanisms become degraded?
[ ] Is there independent monitoring that does not rely on the system under observation?
[ ] Are logs immutable and stored outside the system's control plane?
[ ] Is there a runbook for what to do when monitoring itself fails?

Institutional readiness

[ ] Is stoppability included in the onboarding for every operator?
[ ] Are stoppability drills run on a recurring schedule?
[ ] Is there a post-drill review that updates documentation and thresholds?
[ ] Are stoppability criteria included in procurement and vendor evaluation?
[ ] Is there an executive owner for stoppability who is not the system builder?

Scoring: Count your Yes answers. 20-25: strong stoppability posture. 15-19: gaps that need attention within the quarter. Below 15: stoppability should be prioritized before the next major release or deployment.

For facilitated stoppability drills and governance readiness assessments, schedule a call.

AI Safety Stoppability Checklist | Ethotechnics