Regulators should move toward a strict liability model for AI systems managing critical infrastructure. Under this rule, if a power grid collapses because an AI optimized for cost over stability, the company is liable regardless of intent. This removes the need for plaintiffs to prove specific coding negligence in an opaque "black box" system.
Mandatory telemetry is a practical requirement. AI firms must maintain tamper-proof logs of every optimization decision. When a water plant malfunctions, forensic investigators can trace the specific objective functions that triggered the failure. This mirrors the flight data recorders used in aviation.
Compulsory insurance pools provide another layer of protection. Firms would pay premiums based on the risk profile of their deployment. This ensures the public receives compensation even if the provider goes insolvent. While frameworks like ISO/IEC 42001 offer guidelines, certification shouldn't act as a legal shield. Following a standard does not exempt a corporation from the consequences of a physical catastrophe.
