Definition
Statutory or policy criteria defining when an AI system’s unintended or misaligned conduct is severe enough to trigger emergency government intervention (shutdown, throttle, or capability restriction).
Key Points
- ai-kill-switch-act draft covers unintended conduct causing ≥10 deaths or ≥$100M economic damage (per Ars Technica bill-text summary)
- Also covers goal misalignment, sabotage of shutdown instructions, and concealment from monitoring
- Red-team simulations in controlled environments are typically excepted
- Coverage gated by developer revenue (≥100M) thresholds