Single-tasking when an assistant is one tab away

Single-tasking — one task, fully attended, until a natural stopping point — remains the documented best practice for both performance and wellbeing: cognitive-psychology research has consistently found that multitasking degrades performance on at least one of the interleaved tasks, with task-switching costs measured in time and error rate, and heavy media multitasking associated with poorer attention outcomes in the studies that track it. What AI assistants add is not a new cognitive problem but a new temptation structure: they collapse the cost of starting a side-thought to near zero, so the mental queue of "quick things to check" empties itself into your working session in real time. The defense is structural, not moral.

RechargeMe publishes information about research and practice, not medical or productivity-gospel advice. Studies are named and characterized per their published findings; applications to AI tools are interpretation, labeled as such.

What does the multitasking research actually establish?

Three robust findings. Switching costs: shifting between tasks incurs measurable reconfiguration time — attention doesn't teleport, and the classic experimental literature (Rogers and Monsell's work on task switching, and the decades of follow-ups) quantifies the cost per switch. Interference: concurrent tasks compete for the same processing resources, so at least one degrades — the exception is highly automatized routines. And correlation-heavy but replicated many times: self-reported heavy media multitaskers show, on average, poorer performance on attention-control measures in the studies initiated by Ophir, Nass, and Wagner's 2009 PNAS paper — with the field's honest caveat that causality is contested. The practical consensus underneath the academic noise: fewer switches, better work, less strain.

What's the new AI temptation structure?

Three documented properties of assistants conspire against single-tasking. Zero-friction starts: asking a side question costs seconds, so the threshold for interrupting yourself drops through the floor — this is friction removal, the same force documented for after-hours availability, operating on attention instead of evenings. Parallelizable streams: assistants themselves can run several tasks ("draft this while summarizing that while researching the third"), which invites you to supervise three things instead of doing one — supervision being a euphemism for switching. And conversational momentum: every answered prompt invites a follow-up, so a one-question side trip becomes a ten-turn detour with your original task cooling on the bench. The tool manufactures its own interruptions politely.

Related stories: Late-night chatbot sessions and sleep: what the evidence supports · Does offloading note-taking change how you listen? What the research actually says.

What does the structural defense look like?

Borrow the pattern that already works for notifications, applied to your own queries. The capture list: a paper line or a single note where side-questions go the moment they arise — "check deadline," "ask about vendor terms" — written, not executed; the documented effect of externalizing intrusions is that they stop looping in working memory, the mechanism checklists exploit. Block-shaped AI use: sessions for delegation and retrieval, timeboxed, rather than a continuously open tab — the assistant is a meeting you schedule, not a colleague on your shoulder. One stream per block: when you do delegate, delegate one thing and let it run while you do something unrelated and restorative rather than supervising in a second tab. And the oldest trick in the literature, pre-dating every tool: work in defined units with a stopping point visible — the documented mechanism behind timeboxing and its many branded descendants.

Isn't delegating to AI while you work the whole point?

Partly, and the honest version distinguishes delegation from interruption. Genuine parallelism exists: a long-running generation, a scheduled task, a search that takes minutes — fire it, leave, return. That pattern is fine and is where the tools genuinely add capacity. What the research warns against is supervisory switching — polling the stream, adjusting mid-flight, holding three tasks partially open in your head because each is "almost done." The test is embarrassingly simple: if you're watching it work, you're not delegating; you're multitasking with extra steps. Real delegation has a gap you spend on one other thing — or on nothing, which the wellbeing literature, from attention research to recovery studies, keeps suggesting is the most underrated option on the menu.

What about the wellbeing upside of assistance?

Fairness requires the other column: offloading genuinely draining work — drafting routine email, formatting, first-pass research — can reduce load, and reduced load is the precondition for the recovery the burnout literature documents as protective. The wellbeing question about AI tools is therefore not offload versus no offload; it is whether the reclaimed time and attention actually go to rest or depth, or get refilled with more switches. That outcome is set by structure — the blocks, the capture list, the stopping points — not by the tool's features. The assistant will happily power either result.

FAQ

• Is multitasking actually bad? The experimental literature is consistent: switching costs time and accuracy, and concurrent tasks interfere. What people call multitasking is usually rapid switching, and the bill arrives either way.
• How do I stop interrupting myself with AI side-questions? Capture the question on paper instead of asking it — externalizing intrusions stops them looping — and answer them in a timeboxed AI session later.
• Can't I supervise several AI tasks at once? You can, but supervising is switching. Delegate one long-running thing, leave a gap, return — that's the pattern that adds capacity without the documented switch cost.