Skip to main content

AIPCH21 — Human-Centered Oversight

“Human accountability and intervention controls are proportionate to the AI Product’s agency, embodiment, intelligence effects, risk, and governance obligations.”


What AIPCH21 is really asserting​

AIPCH21 does not require a human to approve every AI action.

It requires the Product to implement the oversight and intervention model that is appropriate to its context and required by the Governance Obligation Profile.

This may include:

  • human-in-the-loop review;
  • human-on-the-loop supervision;
  • approval before action;
  • override/intervention;
  • emergency stop or suspension;
  • escalation;
  • immutable audit and accountability;
  • post-action review.

Positive criteria​

AIPCH21 is Met when:

  1. accountable human roles are explicit;
  2. oversight mode is tied to AR, EM, IE and R0–R4 risk rather than chosen arbitrarily;
  3. mandatory intervention/override/escalation controls are operational;
  4. autonomous action occurs only within delegated and governed boundaries;
  5. audit evidence shows human and AI actions distinctly;
  6. Physical AI has appropriate stop/safe-state mechanisms where required.

Applicability​

Some forms of per-action approval may be Not Applicable for low-risk or explicitly authorized autonomous operation. That does not make AIPCH21 irrelevant: accountability, monitoring, suspension or other oversight mechanisms may still apply.

N/A must be justified from the Product’s Governance Obligation Profile and operating context.


Not Met​

Examples include:

  • required override or intervention control is absent;
  • autonomous physical action lacks required safe-state/stop controls;
  • responsibility for outcomes is unclear;
  • oversight exists only as an informal manual practice;
  • the Product exceeds its delegated agency without escalation.

Evidence signals​

  • Governance Obligation Profile;
  • AR/EM/IE and risk metadata;
  • approval/intervention policies;
  • runtime override/suspension controls;
  • audit trail and escalation evidence;
  • Physical AI safety bindings where applicable.

Agent decision rule​

Met when human accountability and intervention are explicitly governed and proportionate to the Product’s actual authority and effects, without imposing universal HITL where the governance model does not require it.