Habituation

The rigorous version of this is neurobiology, not a design heuristic. Rankin and fourteen co-authors revised the 1966 list of behavioural characteristics in 2007 and defined habituation as a response decrement from repeated stimulation that does not involve sensory adaptation, sensory fatigue or motor fatigue; stimulus specificity and frequency-dependent spontaneous recovery are the two tests that separate it from fatigue. Four of the characteristics run against product intuition: the response recovers at least partially when the stimulus is withheld; more frequent stimulation produces a faster and deeper decrement and faster recovery; within a modality the less intense the stimulus the faster it habituates, and very intense stimuli may show no observable decrement at all; and what restores a decremented response is a different stimulus, which the literature calls dishabituation and which “by itself need not even trigger the response on its own”. The levers that follow are quiet periods, varied presentation and per-stimulus accounting, not volume.

The one interface domain where the mechanism has been measured directly is security warnings, and one research group did all of the measuring. Anderson, Vance, Kirwan, Jenkins and Eargle built a polymorphic warning whose appearance changes on every showing, and under fMRI it was significantly more resistant to habituation than conventional warnings in brain regions related to attention. Vance, Kirwan, Bjornn, Jenkins and Anderson then ran it across a five-day workweek with simultaneous fMRI and eye tracking and found “a dramatic drop in attention throughout the workweek despite partial recovery between workdays”, with the polymorphic design keeping its advantage all five days; their MIS Quarterly paper adds a three-week field study of mobile permission warnings in which adherence fell substantially under standard warnings and stayed high under polymorphic ones. Vary the chrome between showings, and treat the result as one lab replicating itself, because no independent group has reproduced it.

The exposure data that makes alert volume a safety matter comes from hospitals. Drew and colleagues recorded every monitor signal in an intensive care unit over 31 days: 2,558,760 unique alarms, 381,560 of them audible, an audible burden of 187 per bed per day, with 88.8% of 12,671 annotated arrhythmia alarms judged false positives. The Joint Commission's National Patient Safety Goal NPSG.06.01.01 answered with a mandate rather than a tactic: from 1 January 2014 hospitals must “identify the most important alarm signals to manage”, judged partly on “whether specific alarm signals are needed or unnecessarily contribute to alarm noise and alarm fatigue”, and from 1 January 2016 hold policies on settings, on when alarms may be disabled, and on who may set one to “off”. Its rationale cites 98 alarm-related reports in the Commission's sentinel event database, 80 of them ending in patient death, and an FDA database of 566 alarm-related deaths. No controlled test of a tier taxonomy as a product intervention surfaced in this research, so tiering is a regulator-backed practice in one domain, not a measured tactic.

The field tests of the mechanism itself come out negative. Bonafide and colleagues video-recorded nurses responding to 11,745 alarms, 50 of them (0.5%) actionable, at an adjusted median response time of 10.4 minutes — and “the number of nonactionable alarms to which the nurse was exposed in the preceding 120 minutes was not associated with response time”. What did predict response was which alarm it was and how far into the shift the nurse was: a lethal arrhythmia alarm drew 1.2 minutes against 10.4 minutes for other conditions, a one-to-one nursing assignment 3.5 against 10.6 minutes, and each elapsed hour of the shift added 15%. Ancker and colleagues pitted cognitive overload against desensitization in a clinical decision support system, and overload won: reminder acceptance dropped 30% for each additional reminder per encounter and 10% for each five-percentage-point rise in the proportion of repeats, while newly deployed reminders showed no declining response over time. Winters and colleagues' review of 62 articles on clinical alarms found no clear definition of alarm fatigue in use, none measuring it directly, and no agreed valid metric. Instrument repeats per object and per alert type before you attribute a decline to a stale stimulus.

For product notifications the controlled evidence is thinner, and it is about volume. Wohllebe, Hübner, Radtke and Podruzsik randomised 17,500 retail-app users across five push frequencies for seven weeks: as the frequency of non-personalised pushes rose, uninstalls rose and the direct open rate fell, with no significant effect on indirect opens. It is a real field experiment in a minor journal, so weight it accordingly. Fitz, Kushlev, Jagannathan, Lewis, Paliwal and Ariely randomised 237 people to notifications as usual, batched three times a day, or none at all. Batching improved felt attentiveness, productivity, mood, control over the phone, stress and interruptions, while the group that received none reaped few of those benefits and reported higher anxiety and fear of missing out. What you ship against volume is a schedule, not a silence switch.

That resolves the apparent tension between varying an alert and escalating it. Escalation is not naive: characteristic five says that within a modality more intense stimuli habituate more slowly, and very intense ones may not habituate at all. But what restores a decremented response is a different stimulus, not a louder one — Rankin's review cites Marcus, Nolen, Rankin and Carew, who dissociated dishabituation from sensitization in Aplysia by stimulus intensity, with a touch or weak shock dishabituating better than a strong shock, and who concluded that even Groves and Thompson's dual-process account is insufficient. Escalation also carries two costs variation does not. Characteristic nine, habituation of dishabituation, means each escalation round buys less than the last. And louder is measurably worse for the task the user came to do: see banner blindness, where animation improved recall in no condition tested and a target flanked by flashing banners took about 75% longer to find. So “vary when needed” and “do not escalate to beat habituation” are one rule at two grains: varying is the supported move, escalating is the crude version that spends the user's attention on the host task for a diminishing return.

One boundary, and one limit. The downslope of the mere exposure curve — liking rising to about five exposures and back to baseline by fifty — is routinely attributed to habituation, and neither literature demonstrates that: the meta-analysis that established the inverted U concluded that no existing model accounts for it, and the two fields measure different quantities, response magnitude against preference. They also imply opposite things about a change of context. de Zilva, Mitchell and Newell found the preference for familiar targets disrupted when cue-target pairings were rearranged at test, concluding that liking “is specific to the context of exposure”. Habituation is stimulus-specific too, but there the specificity is the good news: change the stimulus and you get the response back. So moving a familiar component to a new screen loses the liking it earned while regaining the attention it lost. Habituation reaches rewards as well as neutral stimuli — Murphy and McSweeney argue that within-session changes in operant responding under a constant rate of reinforcement match the characteristics of habituation and sensitization, which is why a reward stops working once it never varies (see variable reward). Design as though repeated identical alerts lose force, because the exposure data and the frequency experiment support that, but treat “this is habituation” as a hypothesis your own instrumentation has to earn, because the two best field tests of the mechanism did not find it.