What Notifications Actually Do (And What They Don’t)

Less than the usual claims say. An unattended alert measurably increases errors on laboratory attention tasks, and interruptions raise reported stress. But the strongest retrieved trial of turning notifications off found no change in checking, screen time or productivity, and more fear of missing out.

Five different things get filed under “notification overload.” They have five different evidence bases, and they do not agree.

Kept apart: what an alert you never touch does to a laboratory attention task; how long it takes to pick a task back up, measured in a lab; whether real work comes out slower; how the interruption feels; and how much people pick up their phones in the first place. Almost every article on this subject collapses all five into one sentence. The research does not support the collapse, and the most rigorous test in the set points the other way.

An alert you never touch still costs accuracy on a lab task

In a between-subjects experiment, people performing a sustained-attention task made more errors when their phone buzzed or rang, even though they never picked it up and never read anything (Stothart et al. 2015). Two hundred and twelve people were recruited and 166 were analysed. The effect was moderate for text alerts and larger for calls.

Related findings across different laboratory tasks point the same way. A Stroop-task experiment with 105 participants found a notification effect, though not an effect for the mere presence of a phone (Kaminske et al. 2022). An event-related-potential study with 73 participants found small slowing alongside increased neural markers of cognitive control (Upshaw et al. 2022). These are different tasks and different outcome measures, so they are convergent evidence rather than direct replications of Stothart et al.

None of them measures work. The outcome variable is accuracy or latency on a repetitive laboratory task, and nothing in it tells you how a report, a code review or a lesson plan turns out.

Getting back on task takes seconds

The interruption literature has a specific quantity called resumption lag: the time between the end of an interruption and the first response related to the task you were doing. It was measured across a cumulative programme of laboratory experiments totalling 375 participants and 13,377 interruptions, with 25 participants in each between-participants cell (Altmann and Trafton 2007; Altmann and Trafton 2002).

No precise per-position figure is quoted here, because the source reports the lags graphically rather than numerically. What the text does support is a recovery process running over roughly the first 15 seconds after an interruption, and a cumulative cost of 4 to 5 seconds.

This is the point where popular writing goes wrong, and the error is not one of arithmetic. Resumption lag is the latency of the first action on a resumed task, measured at keystroke scale in a laboratory. The office figure people reach for is elapsed time before a person comes back to the task at all, including everything else they did in between. They are not two measurements of the same quantity. One is a within-task latency and the other is a between-task interval, so neither can stand as evidence for the other.

The famous office figure is elapsed time, not recovery time

In an ethnographic study that shadowed 24 information workers for around 700 observation hours, interrupted work was resumed on the same day in 77.2% of cases, on average 25 minutes and 26 seconds later (Mark et al. 2005).

Four things have to be said in the same breath as that number.

It is elapsed time, not recovery time. During those 25 minutes people were working, on an average of 2.26 other working spheres. They were not sitting stunned.

The standard deviation is 54 minutes and 48 seconds, more than twice the mean. A mean with that much spread around it is a poor description of any individual case.

Roughly half of the switches were self-initiated. They were not external interruptions at all. People interrupt themselves.

And the sample is 24 people.

Separately: the recovery figure that circulates most widely in productivity writing, quoted in minutes and seconds as the time it takes to get back to a task, does not come from a peer-reviewed paper. It traces to a 2006 Gallup Business Journal interview (Robison 2006), and the published work from the same research programme reports different numbers. HOS does not use that figure, and neither should anything citing HOS.

In the one controlled test retrieved here, interrupted work came out faster

This is the finding that inverts the standard article. In a 3x2 controlled laboratory experiment with 48 participants, interrupted participants completed the task faster than the uninterrupted baseline: 20.31 and 20.60 minutes against 22.77 (Mark et al. 2008).

Read that carefully in both directions. It does not license “interruptions are good for you.” It is one lab experiment, with 48 people, on one task, and no direct replication of it was located in the documented retrieval for this dossier. What it does is remove the licence for the opposite claim: there is no basis in the retrieved studies for telling readers that notifications generally make real-world work slower.

The cost that showed up was in how the work felt

The same experiment found what it did find. Interrupted participants reported significantly higher stress, frustration, effort, workload and time pressure. Stress moved from 6.92 to 9.46 on a 1 to 20 scale. The authors’ own summary is that interrupted work “may be done faster, but at a price” (Mark et al. 2008).

The subjective direction recurs elsewhere. In a within-subjects crossover with 221 undergraduates, a week of alerts produced higher self-reported inattention and hyperactivity than a week of do-not-disturb, with effect sizes of d=.44 and d=.45 (Kushlev et al. 2016). That study carries a confound that has to be stated with it: the do-not-disturb week also required the phone to be kept out of sight, and the alerts week required it within sight. Phone proximity is confounded with notification state, every outcome is self-report, and the manipulation was not blinded.

The honest summary of this area, and the one HOS will stand behind: some studies report increased stress or subjective disruption, while the strongest retrieved notification-disabling trial was null on most measured outcomes and increased fear of missing out.

Switching notifications off did not change what people did

A preregistered randomised controlled trial with 205 participants, using objective logging rather than self-reported usage, tested disabling notifications (Dekker et al. 2024). It found no effect on checking frequency, screen time, perceived control, overuse, vigilance, productivity or distraction. It did find an increase in fear of missing out.

The authors’ conclusion is that the findings “challenge the assumption that notifications play a prominent role in driving smartphone use” (Dekker et al. 2024). This is the best-designed test of the intervention that every article on this topic recommends, and it came back null on almost everything it measured.

One trial is one trial, but it is preregistered, randomised and objectively logged, which is more than can be said for the intuition it contradicts. The practical question of what to do about your own settings is a separate piece, and this article deliberately does not answer it.

How many notifications a day: the figure in circulation is from 2013

The number people quote is around 65 a day, mean 65.3 and median 63.5 (Pielot et al. 2014). It comes from a mixed-method field study with Android logging and 15 participants, with data collected in February and March 2013, in which social networks accounted for under 4% of notifications. A 15-person sample from early 2013 cannot establish a current population estimate.

The documented retrieval for this dossier, dated 4 September 2026, searching the research databases, did not locate a verified post-2020 device-logged notification-volume figure for adults. Anyone quoting a current daily notification count should be asked where it came from.

The volume is tuned, at the companies that publish about it

At least some platforms publicly describe optimising notifications against engagement objectives, in their own engineering publications. LinkedIn engineers published a system that optimises notification decisions for sessions rather than click-through (Yuan et al. 2022). Pinterest engineers published a notification volume control system that treats daily engagement as “the action to maximize,” on the stated reasoning that optimising for it increases daily active users (Zhao et al. 2018).

That is first-party disclosure rather than inference, and for those two companies it is strong evidence that the number of interruptions reaching you is a tuned parameter rather than a fact of nature. No equivalent disclosure was retrieved for other platforms, and this article does not extend the finding to them.

What this evidence will carry

Keep the five apart and the picture is stable.

Laboratory attention tasks: an unattended alert measurably increases errors. Consistent across the retrieved laboratory studies, which used different tasks, and established for the lab only.

Resumption lag: seconds, in the lab, with a specific operational definition. Established for that construct only.

Real-world work output: not shown to be slower. In the one controlled experiment retrieved here it was faster. No general claim is available in either direction.

Subjective stress and disruption: reported in several designs, with the qualification stated above about what the strongest disabling trial found.

Smartphone checking and total screen time: unchanged by disabling notifications, in the best-designed test retrieved.

The useful thing you can take from this is not a technique. It is a filter. When you next read that interruptions cost workers a fixed number of minutes, you now know which quantity is being described, which one it is being confused with, and why the two cannot be traded for each other. The same filter applies to the wider cost of switching between tasks at work, where the same two quantities are routinely merged.

The wider literature on interruption and focus, including what it does not settle, is set out in the digital distraction and focus guide.

We take one circulating claim at a time and check what the research behind it actually measured. That work will go out in The Weekly System, which has not launched yet. You can join the launch list.