Handing the work to machines and just "keeping an eye on it" sounds like the easy part. Keeping an eye on things is precisely what the human brain does worst.
You're writing a document. Your phone lights up, you glance at it, reply "ok," and come back. The cursor is right where you left it, but how that sentence was supposed to end has evaporated. We tend to think each one costs a few seconds. The real bill isn't in those seconds.
Meanwhile, more and more work is taking a different shape: the machine (an autopilot, an automated pipeline, an AI assistant) does the work, and a person sits beside it, watching and ready to step in. That sounds like liberation from drudgery, but decades of research reach a slightly counterintuitive verdict: long stretches of passive monitoring with the occasional emergency takeover hit several of the brain's weak spots at once. This topic looks at both together, because underneath they run on the same limited "control" machinery.
Psychologists ran a very plain experiment. A letter and a digit appear on screen, say "G7." Sometimes you judge whether the digit is odd or even; sometimes whether the letter is a vowel or a consonant, in a fixed rotation: digits twice, letters twice, digits twice... In 1995 the British psychologists Robert Rogers and Stephen Monsell found that on every trial where the task changed, people were clearly slower and made more errors, even though they knew perfectly well the switch was coming. This is the switch cost.
Where does the time go? Researchers call it task-set reconfiguration. "Judge odd or even" isn't a single thought but a whole bundle of temporary settings: which symbol to look at, which rule to apply, what the left and right keys mean. Switching tasks means unloading the old settings and loading new ones, like clearing the stove before cooking the next dish. In brain imaging, the job falls to a network stretching from the prefrontal cortex to the parietal cortex (toward the top and back of the head). Prefrontal cortex The prefrontal cortex holds the rule for "which task am I doing now," and the parietal cortex helps aim attention at the relevant information.
There's also a gatekeeper: the basal ganglia. Basal ganglia Computational models propose that they decide when a new rule gets "let in" to the prefrontal workbench and when the old one gets kept out. Open the gate too often and every new stimulus pulls you off course; keep it shut too tight and you get stuck in the old task.
The key finding is in the lower half of the figure. If you're told a second or so in advance that the next trial will switch, the cost shrinks noticeably. That part is the workbench you actively set up ahead of time. But the remainder won't go away no matter how much you prepare. It seems the old settings only fully give way once the new task's stimulus actually shows up and you actually do it once. Part of a switch can be prepaid; part of it can only be paid on the spot.
Switch costs are measured in milliseconds. In daily life, a different drag is more expensive. In a 2009 experiment by the American organizational researcher Sophie Leroy, one group was pulled away halfway through a puzzle to do something else; another group finished first and then switched. The people pulled away mid-task did worse on the second task. Part of their mind was still hanging on the first one. She called it attention residue.
Why do unfinished things stick? One explanation: so that you can pick it up later, the brain keeps that goal on standby and periodically pulls it out for a quick review so it isn't forgotten. That review uses up resources that should be going to the new task. A follow-up by Leroy and Theresa Glomb in 2018 found a cheap fix: when interrupted, spend a minute or two writing down "where I am and the first thing I'll do when I come back." Residue dropped noticeably, and performance on the next task improved. Once the goal has somewhere to land, the mind is willing to let go.
Interruptions carry another cost that doesn't show up in output. In a 2008 experiment by the American informatics researcher Gloria Mark and colleagues, frequently interrupted people didn't finish more slowly, and were sometimes even faster, but they reported markedly more stress, frustration and effort. People speed up to make up the time, and the body picks up the tab.
That brings in a hard limit: working memory, the small mental workbench where you temporarily lay out information you need on hand (Topic 3). Researchers estimate it can reliably hold about four items at once. With three unfinished threads open, each takes up a slot, leaving little room for what's in front of you. How many lines you can "watch" at once depends not on willpower but on how big the workbench is and how much unfinished clutter is piled on it.
Large language models have a "workbench" too, called the context window: all the text they can take in at once. When people have AI systems work in relays on long tasks, a common technique is to write a summary or to-do list at each handoff, storing "where we are, what's next," so the next round reads just that note rather than stuffing the whole history back in. It's almost the same idea as the ready-to-resume plan: move the unfinished state out of your head and onto a slip of paper, and the workbench clears. The difference is that models have no residue: switch to a new conversation and the old one is completely gone. Human brains can't be that clean.
If switching is so expensive, why do people still glance at email, messages and progress bars every few minutes? The answer lies in the brain's reward system.
A small group of neurons in the midbrain release dopamine. Reward circuit Dopamine is often called the "pleasure molecule," but more accurately it reports "how much better than expected." The German neuroscientist Wolfram Schultz found in monkeys that unexpected juice makes these neurons fire hard; but once a monkey learns that a certain cue predicts juice, the firing moves to the moment the cue appears, and the juice itself produces no response, because there's no surprise. That gap between "actual minus expected" is called the reward prediction error.
Even more interesting is another 2003 finding from Schultz's group. When juice followed a cue only some of the time, the dopamine neurons showed a slowly rising, sustained firing during the seconds of waiting for the outcome. Its size tracked uncertainty: almost nothing when the reward was certain or certainly absent, and highest at fifty-fifty.
This lines up with what the behaviorist B. F. Skinner saw in pigeons decades earlier: give rewards after an unpredictable number of pecks, and the pigeons work hardest and are hardest to wean off. That's intermittent reinforcement. Checking messages has exactly this structure: most of the time there's nothing, now and then something important or fun arrives, and you never know whether the next check is the one. Each glance is cheap; together they add up to dozens or hundreds of switches a day. To be clear, this links "why we keep checking" to the reward system's computations, but claims like "scrolling = addiction" are far more complicated and can't be read off this directly (Topic 27).
Reinforcement learning is a way of letting machines learn by trial and error, guided by a reward signal. One family of its algorithms, temporal-difference learning, computes at each step a gap of "how much better the actual outcome after this step was than I predicted," and uses it to update the prediction. That gap is the TD error. In 1997 Schultz, Peter Dayan and Read Montague set monkey dopamine firing next to the TD error and found nearly the same shape: an unexpected reward gives a positive error, the response moves to the cue with learning, and a dip appears when an expected reward fails to arrive. It's one of the tightest and best-tested correspondences between neuroscience and AI.
In 1983 the British psychologist Lisanne Bainbridge wrote a paper only a few pages long that has since been cited thousands of times: "Ironies of Automation." Her core point: designers automate everything that can be automated, and what's left for people is exactly the two hardest jobs — watching continuously, and taking over when the machine can't cope.
Why are these two jobs hard? First, the vigilance decrement from the last topic: the duller the watch and the rarer the event, the sooner people flag, with misses climbing within about half an hour. Brainstem arousal system The more reliable the machine, the rarer the "event," and the duller the watch.
Second, automation complacency. In a 1993 flight-simulation study by Raja Parasuraman's group, people handled other tasks manually while monitoring an automated system for failures. When the automation's reliability was consistently high and never changed, people caught its failures noticeably less often than when reliability varied. Interestingly, when monitoring was the only task, people watched very well. Complacency showed up mainly when their hands were busy with other things.
Third, skill decay and being "out of the loop." Skills you don't use slowly erode; worse, you lose track of what the machine is doing right now and why. When it suddenly hands control back, it's usually at the rarest, most complex, most urgent moment. In the 2009 crash of Air France Flight 447, iced-over airspeed probes caused the autopilot to disconnect; the crew suddenly had to fly by hand and, in the confusion, failed to recognize that the aircraft had stalled. The investigation specifically discussed insufficient training in high-altitude manual handling and a loss of situational awareness.
Bring that to today: when one person directs several AI assistants at once, their role shifts from doer to dispatcher and reviewer. Every cost from the earlier sections lands here. Each look at an assistant's output is a switch; each unfinished thread takes a slot on the workbench; and the assistants' rhythm of "sometimes brilliant, sometimes wrong" sits right on that uncertain-reward curve that keeps you checking. The research points not toward "watch harder" but toward redesign: batch reviews instead of being summoned at any moment; attach a "where things stand" note to every interruption; and have people regularly do part of the work by hand, to keep both their skills and their grasp of the system.
Human-in-the-loop means having a person sign off at key points in an automated system, for example an AI assistant asking "OK to proceed?" before a risky action. That design quietly assumes human attention is free and always on call, which is the opposite of everything above. If the assistant pops up a confirmation every few minutes, people soon stop reading and just click approve: automation complacency. If it asks only rarely, the person is out of context by the time it does. So in designing agent orchestration (workflows where several AI assistants divide up a job), "when to interrupt the human, and with how much context" is itself something to design carefully: batch the decisions that need a person, attach a clear summary, and stop only where the risk is real. That's budgeting the human brain's switch cost as a scarce resource.