“It takes 23 minutes and 15 seconds to refocus after an interruption” shows up in productivity apps, management slide decks, and posts about deep work. The figure is real, and it comes from a serious researcher. It does not mean what it is usually taken to mean. None of the peer-reviewed papers on the underlying fieldwork reports 23 minutes and 15 seconds, the fieldwork did not measure focus, and the experiment most often listed as its source found that interrupted people finished their work faster.

The number comes from Gloria Mark, an informatics professor at the University of California, Irvine, and specifically from one interview she gave. Following it back to the original data shows what can be said about interruptions with confidence and what cannot.

One Interview Sentence, Not a Paper

On June 8, 2006, the Gallup Management Journal published a Q&A with Mark about her fieldwork. She described shadowing 36 managers, financial analysts, software developers, engineers, and project leaders for three days each, timing every event to the second. Looking only at work that was interrupted and later resumed the same day, she reported that 81.9% of interrupted work was picked up again that day, after an average of 23 minutes and 15 seconds. She called that interval “not so long.” The bad news, as she framed it, was that people went through about two other tasks before getting back to the original one.

From that interview, the figure spread through business media and productivity blogs, usually without the context. The peer-reviewed version of the research, Mark, González and Harris, “No Task Left Behind? Examining the Nature of Fragmented Work” (ACM CHI 2005), reports a different average from a different number of people.

The two sets of figures are close, but the 15 seconds that make the soundbite feel precise come from a dataset that was never published in full. The paper describes 24 people at a single IT and accounting outsourcing company, observed over 13 months with about 25 hours 42 minutes of timed observation per person. When Mark discussed this work in 2023, around the publication of her book Attention Span, she quoted the paper’s 25 minutes 26 seconds.

What the Stopwatch Was Actually Timing

The unit of analysis was a “working sphere”: a project or task with its own goal, people, tools, and deadline. The researchers analyzed only interruptions that pulled someone out of their current sphere. Interruptions about the same project were left out, and participants often described those as helpful. The clock started when a sphere was abandoned partway through and stopped when the person came back to work in it. It was a time interval between two observable events, not a measurement of a mental state.

The 25-minute average comes with built-in limits. It covers only work resumed the same day, and it excludes interruptions during the last working hour, since those had little chance of being resumed that day. During the gap, people moved through an average of 2.26 other working spheres. Mark has made the same point in recent interviews: for most of those 25 minutes, people were working on other things. The interval shows how long a task sat waiting, not how long a brain needed to get back up to speed.

The time period matters too. The fieldwork happened in the early 2000s, among desk phones, cubicles, and email, before smartphones and team chat. That does not invalidate the result, but it limits how far it generalizes, and the authors themselves say their findings apply only to high-pressure firms similar to the one they studied.

An Average Pulled Up by a Long Tail

The paper’s standard deviation, 54 minutes 48 seconds, is more than twice the mean of 25 minutes 26 seconds. Gaps cannot be negative, so a spread that large points to a strongly right-skewed distribution: many shorter gaps and a minority of very long ones pulling the average up. The paper does not give a median. Our inference is that the typical gap was shorter than the mean, but the data needed to confirm that has not been published.

The breakdowns point the same way. 90.1% of same-day resumptions were self-initiated, and those came back after an average of 21 minutes 28 seconds. The 9.9% that resumed only because a colleague or phone call brought the topic back took 61 minutes 37 seconds. Work interrupted from outside returned sooner (22 minutes 37 seconds) than work people interrupted themselves (29 minutes 1 second), but that difference was only a trend (p = .055). Blog posts presenting it as proof that self-interruptions are worse claim more than the paper shows.

The 23-minute figure is not the time it takes to regain focus. It is the average time an interrupted project waited before someone returned to it, and for most of that time they were doing other work. worldtimedata

The Arithmetic That Breaks the Popular Reading

The usual reading says each interruption costs 23 minutes of lost capacity. Applying that idea to the study’s own numbers shows why it cannot be right.

This is a simplified illustration, since real days included meetings, lunch, and uninterrupted stretches, but the conclusion holds. The people in the study could not have been losing 23 minutes per interruption, because their days were not long enough. The interval overlaps with other work, and separate gaps can run at the same time. Treating it as dead time counts the same minutes several times.

The Most Cited Paper Found Faster Work, Not Slower

Many articles repeating the 23-minute claim cite Mark, Gudith and Klocke, “The Cost of Interrupted Work: More Speed and Stress” (CHI 2008). That paper does not contain the number. It describes a lab experiment co-run with Daniela Gudith and Ulrich Klocke of Humboldt University in Berlin. Forty-eight participants, 81% of them university students with a mean age of 26, played the role of an HR manager answering 12 emails per condition. In the interrupted conditions, a “supervisor” contacted them by phone or instant message about every two minutes.

The measured task time excluded the time spent handling interruptions. Interrupted participants finished sooner: 20.31 minutes with same-topic interruptions and 20.60 minutes with off-topic ones, compared with 22.77 minutes without interruptions. Error rates and politeness scores did not differ significantly. Uninterrupted emails were significantly longer, averaging 31.5 words against 29.2 and 30.2. The cost appeared in self-reports. On the study’s 20-point scale, stress rose from 6.92 to 9.46 (same topic) and 9.13 (off topic), and frustration, time pressure, and effort also increased. The authors concluded that people compensate for interruptions by working faster and writing less, at the price of more strain. They also noted that the effects appeared after only about 20 minutes and said they could not tell whether people would adapt over longer periods.

Other Studies, Other Clocks

An independent field study using a different method found a similar range. Shamsi Iqbal and Eric Horvitz (CHI 2007) installed logging software on the computers of 27 Microsoft employees and recorded 2,267 hours of activity over two weeks. After switching to email in response to an alert, people took an average of 9 minutes 33 seconds to return to any suspended application (8 minutes for instant messages). That was followed by a resumption phase of roughly 11 to 16.5 minutes, depending on the alert type. The authors summarize it as nearly 10 minutes on the alert-driven switch plus another 10 to 15 minutes before focused work resumed. In 27% of cases, people still had not returned to the window they had been working in after two hours.

Lab research measures a different kind of cost: errors rather than minutes. Altmann, Trafton and Hambrick (Journal of Experimental Psychology: General, 2014) gave 300 participants a multi-step procedure that required them to remember where they were in a sequence. Interruptions of about 4.4 seconds roughly tripled sequence errors on the first step after resuming. Interruptions of about 2.8 seconds, the time it took to type two letters, roughly doubled them. This is evidence that very short interruptions are risky in tasks where losing your place leads to mistakes, which is a different claim from “every interruption costs 23 minutes.”

Sophie Leroy’s work on attention residue (Organizational Behavior and Human Decision Processes, 2009) describes part of the mechanism. When people leave a task unfinished, part of their attention stays on it, and performance on the next task suffers. Residue can be measured in experiments, but it is not measured in minutes and cannot be converted into the 23-minute figure.

How to Cite It Without Overstating It

The accurate version is still a strong finding. Most popular versions add claims the data cannot support. These rewordings stay within what the sources show:

What the Data Actually Suggests Doing

The practical lessons in these studies are about resuming work well, not about avoiding a 23-minute penalty. Iqbal and Horvitz found that suspended windows that stayed more than 75% visible were returned to significantly faster than windows mostly covered by others. Windows people had used for 5 to 30 minutes before switching were usually resumed within 5 to 15 minutes of the resumption phase starting, while tasks touched for less than 5 minutes were more likely to go untouched for hours. In Mark’s data, the spheres people returned to on their own were ones they had worked on longer before the interruption (12 minutes 59 seconds versus 8 minutes 58 seconds). Leroy and Glomb (2018), across four studies, found that a brief written plan noting where you stopped and what comes next reduced attention residue when switching to an interrupting task.

These findings point to small, testable habits rather than strict rules. Before switching, write a one-line note on the next step and keep the interrupted task’s window visible. Save protected blocks for tasks where losing your place causes errors, since that is where the lab evidence is strongest. Structures like Pomodoro intervals or time blocking help mainly by moving switches to points you choose, so interruptions land at natural breaks rather than in the middle of a step. That last point is our inference from these studies, not a tested result: none of them evaluated those methods directly.