Twenty-Three Minutes (Recovery Time)

In the AI Story

Mark's research combined time-tracking, observation, and physiological measurement in real office environments, documenting not just when people returned to interrupted tasks but how long it took them to re-engage at the level they'd achieved before interruption. The 23-minute average reflects the central tendency of a distribution with substantial variance, but the floor was revealing: even the fastest resumptions took multiple minutes, and the vast majority took longer than ten. The implication is that quick recoveries from interruption are structurally impossible for complex knowledge work. The cognitive constellation requires time to reassemble, and no amount of individual focus skill or motivational intensity can compress that time below the architectural minimum.

The finding predates AI tools but becomes critically important in their context. Pre-AI knowledge workers were interrupted by meetings, emails, and colleague requests — events occurring on human timescales, often with advance notice, sometimes deferrable. AI agents produce outputs on computational timescales, without advance notice, non-deferrable (the output has arrived and awaits evaluation). The frequency of interruption is therefore higher in AI-augmented work, and the aggregate time lost to recovery correspondingly greater. A builder monitoring three agents who each produce outputs requiring evaluation four times per day experiences twelve interruptions; at 23 minutes each, the recovery cost is 276 minutes — more than four and a half hours — even before accounting for the actual time spent evaluating.

The calculation assumes recovery occurs at all, which task seepage documentation suggests is increasingly false. If the builder uses nominally available time — waits for builds, commutes, lunch breaks — for additional AI interactions, she never experiences the demand-free interval that recovery requires. The resumption lag persists, but it's resumption from a state that was never fully disengaged. The compounding is severe: incomplete recovery produces a higher residue baseline, which produces worse performance on the next task, which produces more errors requiring remediation, which produces more interruptions, which produce more residue. The cycle is self-reinforcing and trends toward progressive degradation unless structural interventions break the loop.

Organizations designed around Mark's finding would look radically different from most contemporary AI-augmented workplaces. They would batch interruptions to minimize their frequency, protect multi-hour windows for uninterrupted work, and treat the 23-minute recovery cost as a real cost to be minimized through workflow design rather than an individual problem to be managed through personal discipline. The scarcity of such organizations reveals a gap between what the cognitive science demonstrates and what the organizational culture is willing to accommodate — a gap that Leroy's simulation argues will determine whether AI-augmented productivity is sustainable or a short-term extraction of human cognitive capital that cannot be replenished at the rate it's being consumed.

Origin

Gloria Mark's finding emerged from a series of studies conducted in the 2000s and 2010s at the University of California, Irvine, using a methodology combining direct observation, computer logging, and physiological measurement (heart rate, stress markers). The 23-minute figure became widely known after Mark's 2008 CHI conference paper and gained mainstream attention through Cal Newport's Deep Work (2016) and Mark's own Attention Span (2023). The precision of the number — 23:15 rather than 'about twenty minutes' — reflects Mark's commitment to empirical exactness and has made the finding memorable and actionable for practitioners and researchers examining the temporal costs of workplace interruption.

Key Ideas

Empirical consistency. The 23-minute average held across multiple studies, diverse work contexts, and different interruption types, suggesting a fundamental cognitive reassembly time rather than a context-dependent variable.

Not just resumption time. The interval measures not when people returned to interrupted tasks but when they regained pre-interruption performance levels — capturing the full cost of cognitive constellation disassembly and reassembly.

Architectural minimum. Even the fastest resumptions took multiple minutes, indicating that rapid recovery from complex-task interruption is structurally impossible regardless of individual focus skill or motivation.

Amplified by AI. AI agents produce more interruptions per day than pre-AI work, and each interruption carries the same 23-minute recovery tax, producing aggregate time costs that can exceed the time spent on productive evaluation.