A model that performs well in the lab may fail on out-of-distribution inputs in real-world use. AI risk management continuously identifies, assesses, and addresses risks to organizations and stakeholders as systems and contexts change. section 1.2
Distribute risk responsibilities across the AI lifecycle:
- Model builders document and communicate known risks and limitations. section 1.2
- Integrators assess how those risks apply to their use case. section 1.2
- Users watch for unexpected behavior during day-to-day operation. section 1.2
Assess technical risks such as bias, drift, and security vulnerabilities alongside societal risks involving fairness, transparency, employment, liberty, and reputation. Use threat modeling to examine design gaps and red teaming to probe vulnerabilities before malicious users do. section 1.2

