AI Safety & Trust and Safety
Human Oversight Policy
Purpose
Defines requirements for meaningful human oversight throughout the AI lifecycle.
Scope
Model development, deployment, and autonomous agent operation.
Definitions
Human-in-the-Loop; Human-on-the-Loop.
Policy Statements
High-risk behaviors and Consequential Actions (per Doc. 20's precise definition) require human-in-the-loop checkpoints enforced at the API level. Autonomous systems support human-on-the-loop dashboards and emergency interrupt controls. Internal governance retains override authority over any deployed model or agent.
Responsibilities
AI Governance Lead owns oversight architecture; Product/Research Engineering implement control interfaces.
Compliance Requirements
Oversight mechanisms validated during Safety Evaluation and Red Teaming.
Enforcement
Systems lacking adequate controls blocked from release.
Reporting Process
Exceptions Process
Reduced oversight for Level 1 features requires Product leadership sign-off.
Review Process
Every 6 months.
Contact Information
Related Policies
This policy references
Referenced by
None