Training Can't Fix Your Brain's Multitasking Bottleneck

Opening

Hello, dear reader. This is Oswarld’s Knowledge Talking.

Today I want to talk about psychology. The word “multitasking” is treated almost like a survival skill for Korean office workers. Checking Slack messages while writing up meeting notes, skimming materials for the next meeting while drafting an email — that’s daily life for many. Many people believe this is possible because of “skill.” And in fact, cognitive psychology¹​ has long held, as its mainstream view, that “with enough practice, two tasks can be processed almost simultaneously.”

But a study published this year directly overturns this 30-year-old belief. The bottom line: no matter how long you train, the brain’s “bottleneck” never disappears. And this finding is more than academic curiosity — it carries significant implications now, at a moment when AI has begun replacing cognitive labor²​.

Twelve Days of Training, and the Collapse of the “Perfect Time-Sharing” Hypothesis

This study was conducted by Professor Torsten Schubert’s team at Martin Luther University Halle-Wittenberg in Germany. It was a joint research effort with FernUniversität in Hagen (Distance University of Hagen) and MSH Medical School Hamburg, and it was published in the Quarterly Journal of Experimental Psychology.

The experimental design is fairly intuitive. Participants were asked to simultaneously perform a visual-manual task (indicating with their right hand the size of a circle briefly shown on screen) and an auditory-verbal task (saying aloud whether a simultaneously heard tone was high, medium, or low). In other words, two tasks that used entirely different sensory channels.

The key lies in the training period. Participants repeated this dual-task³​ for up to 12 days. As a result, reaction speed improved and errors decreased — a pattern consistent with prior research. Earlier studies called this phenomenon “virtually perfect time sharing” and interpreted it as evidence that the brain had achieved true parallel processing⁴​ of the two tasks.

But Schubert’s team took things a step further. After training, they artificially altered the time required for the response selection stage⁵​ of the tasks. In simple terms, they introduced a very small change into a well-trained routine.

The results were dramatic. Lengthening the bottleneck stage of the shorter task also lengthened the reaction time of the longer task — but the reverse did not hold. This is a clear signal that the two tasks were still passing through a single bottleneck sequentially. Practice hadn’t “eliminated” the bottleneck; it had simply made the processing before and after the bottleneck so fast that the bottleneck became invisible.

Professor Schubert explains it this way: the brain is remarkably good at optimizing the sequence of two processes so they don’t interfere with one another. But this optimization has limits, and the moment conditions become even slightly more demanding, the cognitive system quickly tires and errors increase.

The “Latent Bottleneck” Model — Not Gone, Just Hidden

The theoretical framework this study proposes is called the “latent bottleneck model.” Here’s the core idea, unpacked.

The human brain has a central processing bottleneck that two tasks cannot pass through simultaneously. It’s the stage called “response selection” — the process of deciding “what response should I make to this stimulus?” This bottleneck is structural. In other words, it can’t be eliminated by changing strategy or boosting motivation.

Extended practice doesn’t remove the bottleneck itself; instead, it drastically shortens the processing time of the stages before (stimulus recognition) and after (response execution) the bottleneck. As a result, the two tasks appear to pass through the bottleneck almost simultaneously. But if task conditions change even slightly — say, the time required for response selection increases a bit — the hidden bottleneck immediately resurfaces.

Think of it like a highway tollbooth. No matter how optimized traffic flow becomes, the number of lanes at the tollbooth doesn’t change. When traffic is light, you don’t notice the bottleneck — but the moment volume increases even slightly, congestion begins. The brain’s central processing works the same way.

This finding lends weight to the “structural” side of cognitive psychology’s long-running debate over whether the bottleneck is structural or strategic⁶​. It also provides grounds to reinterpret what earlier studies read as evidence of “bottleneck elimination” as, instead, bottleneck “concealment.”

Why This Study Matters Now — Because AI Has Started Taking On Cognitive Labor

Let me shift focus for a moment. This study happens to have been published at a time when, coincidentally, 2025 is the year AI began substantively sharing the cognitive labor of office workers.

McKinsey estimates AI’s long-term productivity growth potential at $4.4 trillion (~₩6,400 trillion). According to ActivTrak’s 2025 State of the Workplace Report, 58% of office workers are already using AI tools at work — a 107% increase from 2022.

But the same report reveals something interesting. Since AI tool adoption, collaboration time rose 27% and multitasking rose 5%, yet focus efficiency⁷​ actually fell from 65% to 62%. Productive time rose 2%, but the average length of focus sessions dropped 8%.

Overlay this data with Schubert’s team’s research, and an uncomfortable picture emerges: the approach of using AI to automate part of a task and then redirecting the “freed-up cognitive resources” toward other work — in other words, AI-assisted multitasking — may not be as effective as we assume.

The expectation that productivity will double if you do other work while AI drafts a report — this rests on structurally the same premise as the hypothesis that “practice enables parallel processing.” But if the latent bottleneck model is correct, our brain’s response selection stage can still only process one decision at a time, no matter how much AI helps.

Professor Tilo Strobach of MSH Medical School Hamburg emphasizes the safety-research implications of this study: “Our results show why multitasking that we think of as having become routine in daily life — like talking on the phone while driving — can be dangerous. The same principle applies to professions that require performing multiple tasks in parallel, such as air traffic controllers or simultaneous interpreters.”

Oz’s Lens

There’s a pattern behind one of the biggest mistakes I keep making. The judgment that “this team is experienced, so they can run two projects at once.” An experienced team really does look like it’s running two projects — reports get delivered, meetings happen, milestones get hit. But at some point — a client makes an unexpected demand, market conditions shift suddenly, a team member drops out — I’ve watched the entire performance collapse over a single small variable. That’s exactly the same structure as what this study calls “the moment the latent bottleneck resurfaces.”

What I’m focused on is the implication for how we design work in the AI era. Many organizations today are designing workflows on the premise that “AI handles repetitive tasks, so people can do more.” But as this study shows, the human cognitive bottleneck is activated by the number of tasks, not their difficulty. No matter how much AI reduces cognitive load⁸​, doing two decisions at once remains impossible.

So here’s how I see it: real productivity gains in the AI era don’t come from “making people do more” but from “making people able to focus more deeply on one thing at a time.” AI’s value doesn’t lie in enabling multitasking — it lies in making multitasking unnecessary.

Closing

Let me sum up. First, no matter how much you train, the brain’s central processing bottleneck never disappears. It only gets faster — it never goes away. Second, the state where multitasking “seems to work well” isn’t the bottleneck being eliminated but the bottleneck being concealed, and it collapses instantly under a small variable. Third, the more AI shares our cognitive labor, the more the core of work design must shift from “more work” to “deeper focus.”

Next time you catch yourself checking a Slack notification while writing a proposal at the same time, remember this: right now, somewhere in your brain, two cars are lined up in front of a tollbooth.

References & Further Reading

• Schubert, T., Liepelt, R., & Strobach, T., “Evidence for a Latent Bottleneck After Extensive Dual-Task Practice of a Visual-Manual and an Auditory-Verbal Task”, Quarterly Journal of Experimental Psychology, 2025. : This is today’s core paper. It contains the results of three experiments and the theoretical basis for the latent bottleneck model.
• Dux, P. E. et al., “A Unified Attentional Bottleneck in the Human Brain”, Proceedings of the National Academy of Sciences, 2011. : A study that used fMRI to identify the brain’s unified attentional bottleneck regions (the inferior frontal junction, superior medial frontal cortex, and bilateral insula). It helps you understand the neuroscientific basis for the bottleneck.
• Strobach, T. & Schubert, T., “A mechanism underlying improved dual-task performance after practice: Reviewing evidence for the memory hypothesis”, Psychonomic Bulletin & Review, 2024. : A paper reviewing the mechanism behind dual-task training effects from a working-memory perspective.
• Pashler, H., “Dual-Task Interference in Simple Tasks: Data and Theory”, Psychological Bulletin, 1994. : A classic of central bottleneck theory. If you want to know where dual-task interference research began, start with this paper.
• ActivTrak, “2025 State of the Workplace Report”, 2025. : An annual report presenting data on changes in focus efficiency, collaboration time, and multitasking among office workers since AI adoption.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. Cognitive Psychology: A branch of psychology that scientifically studies how humans take in information, remember, judge, and act. Think of it as the discipline that uses experiments to uncover “how our brains think.” ↩

2. Cognitive Labor: Work that primarily engages the brain rather than the body. Data analysis, report writing, decision-making, and strategy development are representative examples of this knowledge work. The area AI has begun to replace is a subset of this cognitive labor — summarizing, drafting, pattern recognition, and the like. ↩

3. Dual-Task: An experimental method in which two tasks are performed simultaneously. In everyday life, “talking on the phone while driving” is a classic example. In research, the sensory channels (visual/auditory) or response modes (hand/voice) of the two tasks are deliberately designed to differ, so researchers can precisely measure where interference occurs. ↩

4. Parallel Processing: A method of handling multiple tasks at the same time. In computers, a multi-core CPU can genuinely run several programs simultaneously — but this study’s key finding is that the human brain cannot do this at certain cognitive stages. The opposite concept is serial processing (handling one task at a time). ↩

5. Response Selection Stage: The cognitive process of deciding “what response to make” after recognizing a stimulus. For example, this is the middle of three stages: recognizing with your eyes that a traffic light has turned red (stimulus recognition), deciding “I need to hit the brake” (response selection), and actually moving your foot (response execution). This study identifies this exact stage as the core location of the bottleneck. ↩

6. Structural vs. Strategic Bottleneck Debate: A 30-plus-year academic debate in cognitive psychology over whether dual-task interference arises from the brain’s physical structure (structural) or from a person’s voluntary strategy of processing one task first (strategic). This study’s results support the structural-limitation side. ↩

7. Focus Efficiency: A metric used by ActivTrak representing the proportion of total work time spent absorbed in a single task without interruption. For example, if 5 out of 8 working hours are focus time, focus efficiency is about 63%. The higher this number, the more deep-work time is present. ↩

8. Cognitive Load: The size of the mental burden placed on the brain while performing a task. Easy tasks (driving a familiar route) carry low cognitive load, while difficult tasks (driving to a new place while watching a navigation app) carry high load. AI tools reduce cognitive load by handling repetitive tasks, but according to this study, even when cognitive load decreases, the bottleneck itself — the inability to make two decisions at once — is not resolved. ↩