Human on the loop

Governance

A person supervising a system that acts on its own, intervening by exception rather than by default.

What is a human on the loop?

A human on the loop is a person supervising a system that acts on its own authority. The system completes work without waiting, and the supervisor intervenes when something is flagged or when a review of samples reveals a problem. Intervention is by exception, which makes the quality of the exception signal the whole of the control.

Also called: HOTL

Why the distinction from in the loop is the one vendors blur

Vendors describe both arrangements as human oversight, and the two carry entirely different risk. In the loop, nothing happens until a person acts, so the failure mode is delay. Supervision on the loop lets everything proceed first, with a person finding out afterwards, so the failure mode is a volume of wrong outcomes already delivered. Choosing on the loop is defensible where the action is reversible and the exception signal has been measured; presenting it as equivalent to in the loop is indefensible.

  • Supervision by exception reaches volumes that approval of every outcome cannot.
  • The control rests entirely on what the system chooses to flag.
  • Harm accumulates between an error occurring and a supervisor noticing it.
  • Reversibility of the action decides whether this arrangement is acceptable at all.

How a human on the loop works

  • Actions permitted without prior approval are enumerated explicitly.
  • Confidence thresholds and policy breaches define what raises an exception.
  • The system acts, recording its reasoning for later examination.
  • Flagged cases are presented to the supervisor with the action already taken.
  • A sample of unflagged cases is reviewed independently on a stated cadence.
  • Any pattern found in review narrows the automatic path for the next period.

How a human on the loop is evaluated

This arrangement is judged on how quickly a problem surfaces and on how much it costs before it does. Examine the following over a settled period:

  • proportion of genuine errors that raised an exception
  • elapsed time from an error occurring to a supervisor acting
  • error rate found in sampled unflagged cases
  • volume of outcomes delivered between detection and remedy
  • reversal rate for actions the supervisor later rejected

Limitations, risks and common failure modes

  • Sampling of unflagged cases is abandoned once the deployment appears stable.
  • The exception signal is tuned down because the supervisor is overwhelmed.
  • Irreversible actions are permitted on the assumption that flags are reliable.
  • Supervisors receive flags without the context needed to judge them quickly.

Related terms

Human in the loopA person inside the decision path, whose approval is required before the outcome stands.Escalation policyThe stated conditions under which a system must stop and involve a person.GuardrailsEnforced limits on what a system may do or say, applied outside the model rather than requested of it.RollbackReturning a system to its previous state after a change, which has to be planned rather than improvised.Human approval stepA gate in a workflow where a named person authorises an action before it commits.Jailbreak resistanceHow well a system holds its constraints under deliberate attempts to remove them.

All 315 terms