Global explanations support judging whether to keep relying on the system

What it is

A hospital has to decide whether to keep a sepsis alert next quarter. That decision is not about last night’s one alarm. It is about the apparatus’s temperament on this ward’s patients: how dense the misses are, how noisy the false alarms, on which patients it is worse. The use of a global explanation is to support whether it is still worth relying on.

A crystal-clear account of one instance cannot fill that decision. A well-explained hit does not prove you should stay plugged in next quarter.

Why it happens

Continued reliance is a bet on a future stream of outputs, not a signature on the one that already happened. The bet needs tendency: on the coming distribution, how it will probably behave. Group slices, stability over a window, a list of failure modes — those are the materials for this judgement. One local heatmap tells you why it rang last night; it does not tell you whether the base rate of ringing will wreck the night shift.

Lee and See treat trust as an attitude toward automation that has to line up with reliability. Alignment cannot come from revisiting the most vivid instance; vividness has nothing to do with future frequency. If the global account is missing, the reliance decision spends anecdotes — the last three happened to be right, or yesterday’s one big miss.

Studying it

After seeing a global tendency (misses and false alarms by condition, a three-month trend) or only several locally explained alerts, have a decision-maker choose renew, use in a narrower scope, or switch off. Independent variables: whether the global slice matches this ward’s distribution, whether failure modes are included. Dependent variables: match between the choice and true reliability, how far a single anecdote pulls the choice, whether a narrowed scope lands on the subgroup that is actually worse.

Willingness to renew is not the endpoint. Watch whether the choice lands in the band the data support.

Where it stops holding

A tool someone uses once, in passing, has no “continued reliance” decision; the global account can be omitted. If the system is retrained daily and the distribution drifts fast, a static global account treats expired temperament as evidence that reliance is still safe — more dangerous than none. This entry does not treat how trust recovers after one severe miss; that is a different temporal process.

Applying it

• In confirmations for turning on, renewing, or widening permission, put a global slice: hits, misses, and false alarms in this setting’s recent window, cut by the subgroups the user cares about.
• Do not drive renewal with “look how clearly it explained this one.” Chain that instance back to the trend: typical success, typical failure, or a rare event.
• Offer a narrowed-scope option so the reliance decision can land on a subgroup, not only all-on or all-off.
• Check: with your global panel in hand, ask “will you use it next quarter, and on whom.” If the answer can only restate last night’s alert, the panel is not supporting a reliance judgement.

Related

• Same group: L5.02.1 A local explanation accounts for a single output · L5.02.2 A global explanation accounts for overall behavioural tendency · L5.02.3 Users need different explanation levels in different situations · L5.02.4 Local explanations support accepting or contesting a single result · L5.02.6 An overall impression induced from a few local explanations is often wrong · L5.02.7 Feature importance is a correlation ranking, not a causal account · L5.02.8 Developers need debugging information; end users need grounds for action — they cannot share one explanation
• Nearby: L5.03 Trust Calibration · L5.09 Overtrust and Trust Collapse · L5.01 Types of Explainability
• Search terms: continued reliance · global explanation · trust and reliability