A notification is an interruption, and the evidence says timing beats importance

Year 3 / Module 20 Mobile, ubiquitous and sensor-based computing / Unit 7 / A notification is an interruption, and the evidence says timing beats importance

A notification is an interruption, and the evidence says timing beats importance

HCI3104 · Unit 7 of 10

The interruption literature is the module's strongest evidence and its most counter-intuitive: the cost of a notification depends more on when it arrives than on what it says, so delivering an important message immediately is often worse than deferring it to a natural breakpoint, and the field's engineering effort has moved from suppressing notifications towards predicting the opportune moment to deliver them.

A notification is an interruption, and the evidence says timing beats importance

Iqbal and Bailey (2008) built a system that watches a person work. It finds the boundaries between units of a task, and holds a notification until one arrives. Scheduling to those boundaries cut frustration and reaction time against immediate delivery, at a cost of about a minute and a half of deferral. The result is the strongest evidence in this module and the most counter-intuitive: an important message is not therefore best delivered at once.

This seventh of ten units argues that the field moved from suppressing notifications to choosing when to deliver them. The timing effect has been reproduced with different measures, on desktops and then on real phones. The field has not agreed on what an interruption costs. The units before it made the device sense; this one makes it speak.

Interruption is the normal condition, and its cost is not time

Mark, Gonzalez and Harris (2005) observed 24 information workers and described their work as a sequence of working spheres. It was highly fragmented: people spent little time in a sphere before switching, and 57 per cent of working spheres were interrupted. Most interrupted work resumed the same day, but more than two other activities came first.

What an interruption costs is not elapsed time. Mark, Gudith and Klocke (2008) compared interrupted and uninterrupted work and varied the context of the interruption. Context made no difference. Interrupted participants finished sooner with no loss of quality. They reported more stress, frustration, time pressure and effort, which the authors read as compensation by working faster.

The breakpoint became a design lever, and the policy built on it worked with a qualification

Adamczyk and Bailey (2004) interrupted people at different moments within a task sequence, placing interruptions where a task model predicted cognitive load. Different moments had different effects on emotional state and on attributions about the interrupter. Horvitz and Apacible (2003) then inferred the cost of interrupting a user from device interaction, visual and acoustical analyses and calendar data. They reported how accurately the models predicted it. Interruption becomes an estimate under uncertainty, not a fixed policy: what a scheduler needs.

Iqbal and Bailey (2008) tested the policy itself in two studies. The models detected breakpoints reasonably well but struggled to differentiate their type. Scheduling at breakpoints reduced frustration and reaction time against immediate delivery. The relevance of the content determined which type of breakpoint suited a message.

The lever moved onto the phone itself

Fischer, Greenhalgh and Benford (2011) asked whether opportune moments surface at the endings of episodes of mobile interaction. In a naturalistic experiment, notifications delivered after an episode were attended to significantly more quickly than against a random baseline. The workload of the interrupting task affected delay and completion.

Okoshi and colleagues (2015) built the demonstration. Attelia detects breakpoints from the user's own interaction in real time and delivers at them, without modifying applications or adding dedicated sensors. A controlled study found 46 per cent lower cognitive load than random timing. An in-the-wild study of 30 participants over 16 days found a 33 per cent decrease.

Receptivity can be predicted from the phone, without reading the content

Pielot, de Oliveira, Kwak and Oliver (2014) collected two weeks of behavioural data from 24 mobile instant messaging users. They predicted whether a message would be viewed within minutes. Seven phone features predicted fast attendance with 70.6 per cent accuracy and 81.2 per cent precision. They included notification-centre interaction, screen activity, the proximity sensor and the ringer mode.

Mehrotra and colleagues (2016) combined logging with experience sampling to collect 10,372 in-the-wild notifications and 474 questionnaire responses from 20 users. Response time and perceived disruption depended on presentation, alert type, the sender-recipient relationship and the type, completion level and complexity of the task. A text banner also assumes a reader, and a person reading in a second language is not covered by the timing result. Even important or useful content caused disruption.

The attention is spent anyway, and the drawer holds the rest

Pielot, Church and de Oliveira (2014) logged notifications and diary entries from 15 users for a week. They found 63.5 notifications a day on average, mostly from messengers and email. These were typically viewed within minutes whether or not the phone was silent, with social pressure among the main reasons given.

A larger sample shows where that attention lands when it is not spent. Weber, Voit and Henze (2019) sampled the notification drawers of 3,953 Android devices, collecting more than 8.8 million snapshots. Users held 3.4 pending notifications on average. The count built up overnight and was dealt with in the morning, and messaging sat at the most visible position. The study observes the drawer rather than the person, so it cannot say whether the pending items were wanted or merely unread. It adds the state of a notification delivered well and not attended to, which no timing policy removes on its own.

The contested question: does timing beat content?

The timing position holds that the arrival moment governs the cost. Adamczyk and Bailey (2004) show the moment changing emotional impact. Horvitz and Apacible (2003) show the cost being estimable. Iqbal and Bailey (2008) show scheduling cutting frustration and reaction time. Fischer and colleagues (2011) show activity endings. Okoshi and colleagues (2015) show cognitive load falling by 46 and 33 per cent. Pielot and colleagues (2014) show attentiveness being predictable without content.

The content-and-person position holds that the cost belongs to the message and the relationship. Iqbal and Bailey's second finding is that relevance selects the breakpoint. Mehrotra and colleagues (2016) find important content still disrupting and personality mattering. Pielot, Church and de Oliveira (2014) find social expectation making immediate attention the norm. Pielot and Rello (2017) found 30 volunteers who disabled alerts for 24 hours. They were less distracted and more productive, but also anxious and less connected, so the interruption and the social function arrive together.

The sharpest evidence for the second position comes from outside this unit's list, in the field's own survey. Mehrotra and Musolesi (2017) describe the experiment of Fischer, Yee, Bellotti, Good, Benford and Greenhalgh (2010), which varied content quality and delivery time together. Receptivity was significantly higher for content the participant had rated good, and did not differ significantly between opportune and inopportune delivery.

Iqbal and Bailey (2008) measured reaction time, Okoshi and colleagues (2015) cognitive load, and Pielot and Rello (2017) self-report. Pielot and colleagues (2014) measured the accuracy of a prediction rather than the outcome of an interruption.

What would settle it is a factorial design. It would hold content constant while varying the moment, and hold the moment constant while varying content. It would carry one pre-registered measure through both arms, and run it in the field for weeks with the deferral cost counted. The survey also records that almost all interruptibility studies use small samples for short periods and validate their results offline. It adds that inattentive periods usually last only two to five minutes, which bounds what a deferral can buy.

Rated by kind, the timing side rests on controlled studies with measured outcomes (Adamczyk and Bailey, 2004; Iqbal and Bailey, 2008; Okoshi et al., 2015) and on one field deployment. The content side rests on experience sampling and field logs (Mehrotra et al., 2016; Pielot, Church and de Oliveira, 2014) and on one experiment that varied content and timing together, reported through a survey (Mehrotra and Musolesi, 2017). While the dispute is open, a designer defers low-urgency messages to detected breakpoints and lets content set the urgency. The strongest objection is that content quality, not timing, decided receptivity in the one design that varied both.

Where this is heading

The direction of travel is from scheduling one notification to configuring a channel. Kim and Park (2025) gave users three reception modes, immediate, while in use and on demand, settable per application and per keyword. In a two-week field study with 26 participants, notification volume and habitual phone checking fell, and usage times stayed stable. Okoshi and colleagues (2015) show the mechanism can run on the phone without touching applications, which makes it a platform feature rather than a prototype. What has not changed since Adamczyk and Bailey (2004) is the unit of the decision. When a message arrives is a design variable, and the argument is over how much of the outcome it governs.

Sources

Adamczyk, P. D. and Bailey, B. P. (2004). If not now, when? The effects of interruption at different moments within task execution. CHI. doi:10.1145/985692.985727 Interrupted people at different moments within a task sequence, placing interruptions where a task model predicted cognitive load. Different moments had different effects on emotional state and on attributions about the interrupter, fixing the arrival moment as the design decision.
Fischer, J. E., Greenhalgh, C. and Benford, S. (2011). Investigating episodes of mobile phone activity as indicators of opportune moments to deliver notifications. MobileHCI. doi:10.1145/2037373.2037402 Tested whether opportune moments surface at the endings of mobile interaction episodes such as a call or an SMS, against a random baseline. Notifications arriving after an episode were attended to significantly more quickly, though the workload of the interrupting task affected delay and completion.
Horvitz, E. and Apacible, J. (2003). Learning and reasoning about interruption. ICMI. doi:10.1145/958432.958440 Inferred the cost of interrupting a user from device interaction, visual and acoustical analyses and calendar data, and tested how accurately the models predicted it. Interruption cost becomes an estimate a system can compute under uncertainty, the precondition for scheduling.
Iqbal, S. T. and Bailey, B. P. (2008). Effects of intelligent notification management on users and their tasks. CHI. doi:10.1145/1357054.1357070 Built a notification management system that deferred notifications to detected breakpoints and evaluated it in two studies. The models detected breakpoints reasonably well but struggled to differentiate their type; scheduling at breakpoints cut frustration and reaction time against immediate delivery; and content relevance determined which breakpoint suited a message.
Kim, J. and Park, S. (2025). Empowering individual preferences in mobile notifications: a balanced approach to cognitive load and information needs. IEEE Access. doi:10.1109/access.2025.3549033 Proposed user-driven reception modes, immediate, while in use and on demand, and evaluated them in a two-week field study with 26 participants. Notification volume and habitual phone checking fell and participants reported less notification fatigue.
Mark, G., Gonzalez, V. M. and Harris, J. (2005). No task left behind? Examining the nature of fragmented work. CHI. doi:10.1145/1054972.1055017 Observed 24 information workers and analysed their work as a sequence of working spheres. Little time was spent in a sphere before switching, 57 per cent of working spheres were interrupted, and most interrupted work resumed the same day but only after more than two other activities.
Mark, G., Gudith, D. and Klocke, U. (2008). The cost of interrupted work: more speed and stress. CHI. doi:10.1145/1357054.1357072 Compared interrupted and uninterrupted work while varying the context of the interruption. Context made no difference; interrupted participants finished sooner without loss of quality but reported more stress, frustration, time pressure and effort, so the cost sits in the person's state rather than in the time taken.
Mehrotra, A., Pejovic, V., Vermeulen, J., Hendley, R. and Musolesi, M. (2016). My Phone and Me: understanding people's receptivity to mobile notifications. CHI. doi:10.1145/2858036.2858566 Collected 10,372 in-the-wild notifications and 474 questionnaire responses from 20 users, combining logging with experience sampling. Disruption depended on presentation, alert type, the sender-recipient relationship and the task, and even important content caused disruption.
Mehrotra, A. and Musolesi, M. (2017). Intelligent notification systems: a survey of the state of the art and research challenges. arXiv:1711.10171. https://arxiv.org/abs/1711.10171 Surveyed interruptibility and receptivity research and set out the open challenges. It reports Fischer, Yee, Bellotti, Good, Benford and Greenhalgh (2010), where receptivity was higher for good content but did not differ between opportune and inopportune delivery, and records that studies use small samples for short periods.
Fischer, J. E., Yee, N., Bellotti, V., Good, N., Benford, S. and Greenhalgh, C. (2010). Effects of content and time of delivery on receptivity to mobile interruptions. MobileHCI. doi:10.1145/1851600.1851620 Varied content quality and delivery time together and found receptivity higher for good content with no difference between opportune and inopportune delivery. It is the comparison the survey above reports, and it was run in one laboratory setting.
Okoshi, T., Ramos, J., Nozaki, H., Nakazawa, J., Dey, A. and Tokuda, H. (2015). Attelia: reducing user's cognitive load due to interruptive notifications on smart phones. IEEE PerCom. doi:10.1109/percom.2015.7146515 Presented Attelia, middleware that detects breakpoints in phone interaction in real time and delivers notifications at them, without modifying applications. A controlled study found 46 per cent lower cognitive load than random timing, and a field study of 30 people found 33 per cent.
Pielot, M., Church, K. and de Oliveira, R. (2014). An in-situ study of mobile phone notifications. MobileHCI. doi:10.1145/2628363.2628364 Logged notifications and diary entries from 15 mobile phone users for a week. Participants dealt with 63.5 notifications a day, mostly from messengers and email, and typically viewed them within minutes whether or not the phone was silent, with social pressure among the reasons given.
Pielot, M., de Oliveira, R., Kwak, H. and Oliver, N. (2014). Didn't you see my message? Predicting attentiveness to mobile instant messages. CHI. doi:10.1145/2556288.2556973 Collected two weeks of data from 24 mobile instant messaging users and predicted whether a message would be viewed within minutes. Seven phone features predicted fast attendance with 70.6 per cent accuracy, while the availability cues apps expose were weak predictors.
Pielot, M. and Rello, L. (2017). Productive, anxious, lonely: 24 hours without push notifications. MobileHCI. doi:10.1145/3098279.3098526 Ran the Do Not Disturb Challenge, in which 30 volunteers disabled notification alerts for 24 hours across all devices. Participants were less distracted and more productive but also anxious and less connected, and about two thirds intended to change how they manage notifications.
Weber, D., Voit, A. and Henze, N. (2019). Clear All: A Large-Scale Observational Study on Mobile Notification Drawers. Mensch und Computer. doi:10.1145/3340764.3340765 Sampled the notification drawers of 3,953 Android devices, collecting more than 8.8 million snapshots. Users held 3.4 pending notifications on average, and the count built up overnight.

Check your understanding & apply it · 4 questions, 1 essay

1A team ships a smart delivery feature that holds every notification, including personal messages, until the phone detects a breakpoint in interaction. Diagnosing from this unit, which users does the feature serve worst and why?

2A product increases how often it sends notifications, on the assumption that important content will always be attended to. Predicting from this unit, what happens?

3A vendor cites Attelia to promise that its own users will see a 33 per cent cut in cognitive load from breakpoint delivery. Diagnosing from this unit, which option is correct?

4Two reviewers disagree. One says the arrival moment governs the cost of a notification; the other says the message and the relationship do. Adjudicating from this unit, which option is correct?

5Essay A team is preparing to ship a messaging app with a smart delivery feature that holds notifications back until the phone detects a breakpoint in the user's interaction, on the argument that timing matters more than the content of a message. Advise the team on what this unit's evidence does and does not support, which users the feature will serve worst, how the deferred messages should be treated while they wait, and how the team should measure and report the result before release. Cite the sources you rely on at each step.

The evidence supports the mechanism. Iqbal and Bailey (2008) found that deferring notifications to detected breakpoints reduced frustration and reaction time against immediate delivery, at a cost of about a minute and a half, although their models detected breakpoints reasonably well while struggling to differentiate their type; Fischer, Greenhalgh and Benford (2011) found that notifications delivered after an episode of mobile interaction were attended to significantly more quickly than against a random baseline; and Okoshi and colleagues (2015) found 46 per cent lower cognitive load at detected breakpoints than at random timing, and a 33 per cent decrease in a 16-day in-the-wild study with 30 participants. Horvitz and Apacible (2003) supply the framing: the cost of interrupting is an estimate under uncertainty, which is what the system is computing. The evidence does not support the claim that content can be ignored. Iqbal and Bailey's own finding is that content relevance determines which breakpoint suits a message. Mehrotra and colleagues (2016) found that even a notification with important or useful content caused disruption and that presentation, alert type, sender-recipient relationship, task complexity and individual psychological traits all moved the outcome. The feature will serve worst the users whose reason for checking the phone is social: Pielot, Church and de Oliveira (2014) found 63.5 notifications a day typically viewed within minutes whether or not the phone was silent, with social pressure among the main reasons. Pielot and Rello (2017) found that 30 volunteers who disabled alerts for 24 hours were less distracted and more productive but also anxious about being less responsive and less connected. The team should therefore treat deferral as a social decision, hold personal messages briefly and visibly rather than silently, offer per-application and per-keyword control of the kind Kim and Park (2025) evaluated, and report what a deferred message cannot do: recover the social cost. Measurement should follow Mark, Gudith and Klocke (2008), who found interrupted work finished sooner without loss of quality but with more stress, frustration, time pressure and effort, so completion time alone will understate the cost; and Mark, Gonzalez and Harris (2005), who found 57 per cent of working spheres interrupted with more than two other activities intervening before resumption. The team should also answer the survey's criticism, reported by Mehrotra and Musolesi (2017), that studies use small samples for short periods and are validated offline, by reporting a field trial with a stated sample, its duration, and the share of messages whose deferral exceeded the two to five minutes that inattentive periods typically last.