Part three

Trust, explanation and failure

The operator is a safeguard only while their judgement stays independent of the system they are safeguarding. Calibration is the discipline that keeps it so.

01Psychology of fallibility

Automation bias and anchoring

The operator is a safeguard only while their judgement remains independent of the system they are safeguarding.

Under-trust · needless overrideCalibrated relianceOver-trust · automation bias
Performance peaks where reliance matches demonstrated reliability. Both tails cost missions — one in wasted operator capacity, the other in unexamined error.

Automation bias

Operators systematically over-weight machine-generated recommendations, both by accepting incorrect output (commission) and by failing to act on a problem the system did not flag (omission). Bias grows with time pressure, workload and the perceived sophistication of the interface.

02Interpretability

Operationalising explainable AI

Explanation is an operational instrument, not a compliance artefact. It has to be readable inside the decision window it belongs to.

Reading saliency and attribution

Visual saliency maps and feature attributions indicate where a model's output was driven from. They tell you what the model attended to, not whether attending to it was correct. A saliency map centred on a legitimate feature is weak evidence; a map centred on an artefact of capture is strong evidence against the output.

03Safeguards

Designing against over-reliance

Failure-mode safeguards are procedural as much as technical, and they are what makes an SLO's output defensible.

Check any term in the glossary, then run a mission.