Long stretches without failure lower monitoring intensity

What it is

When a system runs itself well for a long time, how often and how deeply people look at it drops. Parasuraman, Molloy and Singh called this slackening of watch on reliable automation automation complacency. It is not following a bad suggestion once. It is scanning, checking and standing by being dropped over a quiet stretch.

The gauges still refresh. The eyes are no longer on the gauges.

Why it happens

Monitoring is a standing cost. Reliable automation turns anomalies into very low-probability events; the immediate return on each scan approaches zero. Under dual task, people move resources to the task with immediate feedback, and the automated channel is left with an occasional glance. What Wiener saw in the cockpit is this reallocation of resource, not arrogance of attitude.

Complacency needs time to accumulate. People on a first shift often over-watch; complacency appears after “it has never failed on me” becomes personal experience. If the UI renders quiet as success (all green, no notice means normal), the design itself confirms that not looking is fine.

Studying it

Dual task: a primary task that needs continuous control, and a gauge or log watched by automation, with occasional faults. Independent variables: length of the reliable stretch before a fault, presence of a second task, salience of the fault signal. Dependent variables: time to detect the fault, miss rate, gaze time or sample count on the automated channel.

In a single task people can stare the whole time and complacency is hard to produce; a lab without a second task will understate the problem. Subjective “I was monitoring” must be logged separately from actual sampling.

Where it stops holding

When a system is new, faults are frequent, or operators are scored on “confirm every minute,” monitoring intensity will not drop by itself — the cost is load and ritual. Fully manual work has no such curve. This entry only treats how quiet time wears monitoring down; not whether the rare fault then lands on the weakest moment, nor how to pull the person back into the loop on purpose.

Applying it

• Do not let “no notice” be the only normal. For an automated channel that can sit quiet for a long time, require a brief active sample on a schedule (open the detail, confirm the current value). The sample must meet the content.
• In a dual-task layout, the automated channel cannot signal health only by turning green; health must cost one beat of attention, not zero.
• Check: let the system run quiet for the stretch you currently treat as “safe to relax,” and count real samples of the automated channel in that stretch. If the count is near zero, complacency is already in the product; you do not need to wait for a fault.

Related

• Same group: L4.03.2 Rare failures land exactly when monitoring is weakest · L4.03.3 Human engagement has to be actively maintained
• Nearby: L4.02 Automation Bias · L4.09 Skill Degradation · L5.09 Overtrust and Trust Collapse
• Search terms: automation complacency · monitoring of automation · out-of-the-loop