False alarms teach people to ignore real ones

What it is

Alarm fatigue is not "it rings too often." It is low positive predictive value: it rang, and there was usually no danger to handle. After a few rounds, the judgment of the next ring moves from "might be something" to "probably nothing." Real alarms then get the response reserved for false ones—delay, mute, skip. Clinical monitors, security, ops, and in-car systems have all eaten this. In consumer notices, calling every ordinary sign-in an "anomaly" trains the same reaction.

This is how false positives poison true alarms. It is not labeling every item top-tier (that is scale collapse), and not interruption cost rising with engagement (that is timing and depth).

Why it happens

An alarm is a hypothesis test: sound = the system claims a state that needs intervention. People update the credibility of that claim with recent hit rate. False alarms pull the posterior toward "claim is false." Attention is allocated by credibility; when the posterior is low enough, even detection is skipped—not unheard, but heard without launching a handling program.

The danger is that true and false share one sound, one place, one sender. People cannot tell them apart before handling starts, so distrust of the channel lands on every arrival at once. Making the true alarm louder changes detection. It does not change the learning that "after I handle it, it was false again."

Studying it

In a task with a checkable true state (a simulated monitor, a dashboard with injected faults), manipulate false-alarm rate and record response time to true events, misses, and silencing or threshold-changing behavior.

Independent variables: false-positive rate, baseline rate of true events, whether true and false share sound and color. Dependent variables: misses and delay on true events, share of alarms muted, later ability to recall the last true event.

Labs that highlight true events will understate misses. Harder: place a true event after a long false sequence with no warning that a real one is coming. Satisfaction is not a substitute for misses—people can find alarms annoying and still catch the true ones, or the reverse.

Where it stops holding

Marketing pushes with no checkable "true state" are annoyance and disable, not alarm fatigue; using that data as fatigue evidence mixes the constructs. Extremely rare true events (a fire once a year) can leave people who have never seen a true positive even when the absolute false rate is modest—the channel will keep feeling false; drills, not daily false rings, have to carry credibility. When automation already handles the response, human ignoring may not cause a miss—until the day automation fails.

Applying it

• Alarm only for states whose non-handling causes irreversible loss; fluctuations that belong on a status page should not be called alarms.
• If detection cannot yet separate true from false, at least separate presentation: true alarms get their own sound and place, not the everyday cue slot.
• Every alarm should be traceable as true or false after the fact. Falses go into a threshold-tuning queue, not into another identical ring.
• Verify by labeling a week's sends as later-confirmed true or false. If falses dominate and true items are not answered faster than falses, the channel is already written by error. Stop the falses first, then see whether true response returns.

Related

• Within the group: H5.08.2 Once fatigue sets in, it is hard to reverse · H5.08.3 Cut false alarms before turning the volume up
• Adjacent: H5.01 Urgency grading · H5.02 Interruption cost and timing · H3.02 Three elements of error messages
• Search terms: alarm fatigue · false positive · positive predictive value