High-risk AI system checklist
High-impact AI systems require stronger evidence because errors can affect rights, access, safety, or essential services.
Impact profile
Identify affected groups, decision stakes, autonomy level, reversibility, and appeal options.
Control design
Define human oversight, monitoring, fallback, change control, and incident response.
Validation depth
Test performance, bias, robustness, security, and failure behavior in realistic operating conditions.
- Use stricter launch gates when harm is hard to reverse.
- Document residual risk and approval authority.
- Review legal obligations separately.