Attention and Concentration

David Park writes code for a living. Complex algorithms, thorny bugs, systems whose architecture he has to hold in his head all at once. In 2023, the 28-year-old engineer could sit with a single problem for the better part of a morning — tracing an execution path, keeping a dozen moving parts suspended in working memory, the whole edifice slowly resolving into something he understood. It was demanding. It was also the best part of the job.

By 2025 the mornings felt different. He would open his editor and, within minutes, his attention would splinter. A glance at the documentation, a jump to Stack Overflow, a query to his AI coding assistant, a flick over to Slack, back to the editor to evaluate whatever the assistant had suggested. The assistant genuinely helped — it autocompleted functions, drafted boilerplate, proposed fixes he'd have spent twenty minutes writing himself. And yet each suggestion tugged at him. Each one had to be read, judged, accepted or thrown away. He was shipping as much code as ever, maybe more. But he almost never dropped into the deep, absorbed concentration that used to define his best work. He was busier and shallower at the same time.

David's account — a composite of a pattern now reported across writing, design, research, and analysis — captures the puzzle this chapter is about. The tool built to make him more productive seems also to be fragmenting the very faculty his most valuable work depends on. But to understand what is actually happening, we have to start somewhere uncomfortable: with the numbers everyone quotes, and how few of them survive contact with scrutiny.

The Number That Isn't What You Think

You have almost certainly encountered the claim. The human attention span has collapsed from 12 seconds in 2000 to around 8 seconds today — shorter, the punchline goes, than a goldfish's nine. It appears in TED talks, consulting decks, and news segments. In the framing this chapter was asked to examine, the trajectory continues: 12 seconds in 2000, 9.2 seconds in 2022, 8.25 seconds in 2025.

Here is what we know about those figures: they do not measure what they claim to measure, and in the strict sense they do not measure anything at all.

The lineage is traceable. The "8-second" statistic entered wide circulation through a 2015 Microsoft Canada consumer report on advertising, which attributed it to a data-aggregation site called Statistic Brain. When the BBC journalist Simon Maybin tried to run the claim to ground in 2017, the trail dissolved. Statistic Brain's cited sources either could not be located or, when contacted, said they had never produced such a figure. The goldfish comparison had no identifiable scientific origin whatsoever. Neuroscientists who study attention for a living told the BBC plainly that the idea of a single, species-wide "attention span" that could be clocked at eight seconds and tracked year over year is not how attention works or how anyone measures it. There is no longitudinal instrument sitting somewhere, taking humanity's attentional pulse annually and watching it fall by tenths of a second. The precision of "9.2" and "8.25" is exactly the tell: real measurement of a fuzzy human capacity does not arrive pre-rounded to two decimal places from an unnamed source.

So the honest answer to the chapter's central question — what do these figures represent and how are they measured — is deflating but important: they represent a viral statistic with no methodology behind it, and they are not measured at all. Any account of AI's effect on attention that leans on them is building on sand.

This matters beyond pedantry, because there is a real, carefully measured decline in attention — it just looks different, and the difference changes the whole story.

The rigorous work belongs to Gloria Mark, an informatics researcher at the University of California, Irvine, who has spent two decades doing something the goldfish statistic never did: actually observing people. Using screen-logging software and direct observation of knowledge workers at their desks, Mark and her colleagues measured how long a person stays on a single screen before switching to another. In 2004, the average was about two and a half minutes. By around 2012 it had fallen to roughly 75 seconds. In her more recent studies it sits near 47 seconds. That is a real, replicated, methodologically grounded finding — and notice what it is a finding about. It is not a measure of the maximum span a human can sustain when they want to. It is a measure of how frequently, in a working environment saturated with screens and prompts, they in fact switch. It describes an environment acting on people, not a biological capacity that has shrunk.

The distinction is the hinge of this entire subject. The 47-second figure tells us the conditions for sustained attention have degraded, not that the human ability to concentrate has been neurologically amputated. That reframing — from a broken faculty to a hostile environment — is what lets us ask the right question about AI's specific role.

Untangling AI From Everything Else

Which brings us to the hardest problem in the chapter, and it is a problem of evidence rather than of technology. Mark's 47-second decline runs from 2004 to the present. For most of that span, there was no AI in the picture at all. The forces that drove the number down were the smartphone, the always-on browser tab, the notification, and above all the algorithmically optimized social feed engineered to be infinitely scrollable. By the time generative AI tools arrived in workplaces at scale — late 2022 onward — the attentional environment was already deeply fragmented. AI landed on ground that smartphones and social media had spent fifteen years tilling.

This makes isolating AI's specific contribution genuinely difficult, and intellectual honesty demands we say so. In any real worker's day, the smartphone, the social feed, and the AI assistant are all present at once. If David Park's focus is worse in 2025 than in 2023, how much of that is the coding assistant and how much is that he is also two years deeper into a phone habit? Cleanly separating one digital influence from the others is the central methodological challenge, and most of the confident claims you will read about "AI shrinking your attention" have not solved it.

Researchers have a few ways to try. The strongest is the controlled comparison: take a task, have one group do it with an AI tool and a matched group do it without, hold everything else constant, and measure the difference. This isolates the tool's effect on that task in that moment. A second approach is within-person tracking — following the same workers as AI adoption rises in their workflow and watching whether their attention patterns shift in step. A third is neurophysiological: putting people in front of an AI task and measuring the brain directly, so the outcome doesn't depend on self-report. Each method buys some causal traction. None fully escapes the fact that the broader digital environment keeps changing underneath the study.

There is also a direction-of-causation trap that deserves naming. Suppose we find that heavy AI users have more fragmented attention. That does not establish that AI fragmented it. It is entirely plausible that people who already struggle to sustain focus are the ones most drawn to tools that offer instant answers and eliminate the slow slog of doing it themselves — that pre-existing distractibility drives adoption rather than the other way around. To claim AI causes fragmentation, researchers need evidence that AI use came first — either through experiments where they assign the tool, or through longitudinal designs that establish the timeline. Much of the current literature, honestly assessed, cannot yet rule out the reverse story. So the correct posture is this: we have solid evidence that the digital environment as a whole has fragmented attention, suggestive experimental evidence that AI tools impose real cognitive costs in the moment, and genuinely thin evidence that AI specifically is driving a durable, brain-level decline distinct from what smartphones already did. The rest of this chapter tries to hold that calibration steady even where the effects it describes are real.

When Assistance Is the Distraction

The interesting claim about AI is not that it distracts you the way a buzzing phone does. It is that AI fragments attention through the act of helping — that the very features making it useful are the ones dividing your focus. This is what separates an AI assistant from a passive tool. A word processor waits. A spreadsheet waits. They sit inert until you act. An AI assistant is an active participant that reaches back, and that difference runs through four distinct mechanisms.

The first is the autocomplete suggestion. When David codes, his assistant surfaces completions every few seconds, and each one forces a micro-decision: accept, reject, or modify? Individually trivial. Collectively, they shift his attention off the problem he is solving and onto the stream of the tool's proposals. He is no longer only thinking; he is also continuously supervising something that thinks alongside him, and supervision is its own cognitive job.

The second is the frictionlessness of context switching. Generating a draft, querying a knowledge base, producing a snippet — each is nearly instantaneous, so nothing discourages jumping between them. The old friction that used to hold you in one task long enough to go deep (it was simply too much effort to switch) has been engineered away, and with it the incidental discipline it imposed.

The third is iterative prompting. Getting good output usually means a loop: ask, read, refine, re-ask. That loop is engaging by design, and it keeps you bound to the conversation with the tool rather than to the underlying problem the tool is supposed to help you solve. You can spend twenty minutes negotiating with an assistant about a paragraph you could have written in ten, and feel busy the whole time.

The fourth is evaluation overhead, and it may be the heaviest. AI output cannot be trusted blindly, so it has to be checked. Is this code correct? Is this fact real? Is this argument coherent? Verification is a genuine cognitive load, and it competes directly with the focused reasoning the task itself demands. Worse, evaluating someone else's proposed answer is a different mental mode than generating your own — more reactive, more shallow — and the more of your day is spent judging AI drafts, the less of it is spent in the generative depth where the hardest thinking happens.

None of these are malfunctions. They are the product working as intended. Interactivity and sustained absorption are simply at odds: the more responsive and engaging a tool is, the harder it is to disappear into the uninterrupted state that deep work requires.

What the Brain Studies Actually Show

The most direct evidence comes from watching the brain while it works, and here one study has dominated the conversation. In 2025, a team at the MIT Media Lab led by Nataliya Kosmyna ran an EEG experiment — "Your Brain on ChatGPT" — in which participants wrote essays under three conditions: using a large language model, using a conventional search engine, or using only their own heads. The brain-only writers showed the strongest and most widely distributed neural connectivity. The LLM-assisted writers showed the weakest, with notably reduced connectivity in the frequency bands associated with focused, integrative thinking, including reduced alpha-band activity. Alpha waves are a signature of relaxed, internally directed concentration — the neural fingerprint of a mind comfortably absorbed. The LLM group's brains looked, instead, less internally engaged and more oriented toward the external tool. Strikingly, when asked shortly afterward to quote from the essays they had just "written," many LLM users couldn't. They had produced text without encoding it — output without ownership.

A parallel strand of interpretation concerns the prefrontal cortex, the seat of effortful executive control. The suggestive pattern is that AI-assisted work keeps the prefrontal cortex in an active, decision-making state even during tasks that experience should have made fluent and automatic. Normally, practice lets a skill sink below conscious effort, freeing executive resources for higher-order thought — the way a fluent reader stops sounding out words and simply reads. If AI keeps pulling the brain back into deliberate, choice-by-choice supervision, it may block that settling, leaving routine work perpetually effortful.

The honest reading of these findings requires care, because they are easy to over-sell. The MIT study is a small sample and, at the time it swept through the press, a preprint that had not completed peer review. It shows a real difference in brain state while using the tool — which is important and plausible. What it does not show, and what no study yet cleanly shows, is that these in-the-moment differences accumulate into a permanent rewiring of the brain's baseline. The leap from "your brain looks different while leaning on ChatGPT" to "AI has changed what your brain does when the tool is gone" is exactly the leap the evidence does not yet license. There are hints — workers describe their attention continuing to fracture even when they deliberately work unassisted, as though they have trained themselves to expect an interruption that never comes and find its absence unsettling rather than restful. That is a real and widely reported experience. But reported experience is not the same as demonstrated neural change, and we should hold the "conditioning persists after the tool is gone" claim as a serious hypothesis, not an established fact.

The Interruption Arithmetic

AI does not act on a calm mind. It acts on one already under siege. A typical knowledge worker's day is stitched through with alerts — colleague messages, review requests, document comments, calendar pings, build-status updates — and now the suggestion streams of AI tools layered on top. Each is small. Each is, in isolation, helpful. Together they build an environment in which sustained focus is structurally, not merely personally, hard to reach.

The cost of this is where Gloria Mark's other famous number earns its keep. Her research on interrupted work found that after a distraction, it takes an average of about 23 minutes to return to the original task at the original depth. Sit with the arithmetic. If genuine recovery takes 23 minutes but interruptions arrive every few minutes — as they routinely do in a notification-saturated workplace — then the worker never actually completes the return trip. They are perpetually climbing back toward a depth they never reach before the next interruption knocks them down again. Deep focus doesn't just become difficult; it becomes mathematically unavailable.

And this is precisely where AI tools are structurally distinct from the passive clutter around them. An inbox waits to be checked. A document sits until opened. AI systems, by contrast, initiate — they offer, they prompt, they generate things that demand a response. They are on the active side of the ledger, adding to the interruption count rather than sitting quietly among the things you might choose to check. The companies building them optimize, understandably, for engagement: how often you interact, how long you stay, how frequently you come back. Engagement is measurable and monetizable. The thing that engagement crowds out — long, unbroken stretches of concentration — is invisible in the analytics, and a stretch of non-use can even be misread as underutilization of the product. The incentive of the tool and the attentional need of its user point in opposite directions.

The Generational Question

The fragmentation is not evenly felt, and the difference is instructive — though here, too, some caution about the specific figures is warranted. The framing this chapter examines cites a survey finding that 68% of workers under 30 report difficulty maintaining focus for more than twenty minutes without checking a device or tool, against 34% of workers over 50. Take the exact percentages as illustrative rather than gospel — the underlying survey and its methods are not well established, and self-reported focus difficulty is a soft measure. But the direction is consistent with everything the rigorous research implies, and the direction is what matters.

The plausible reading is not that young workers have weaker willpower. It is that they were conditioned earlier. Someone who came of age inside the smartphone's attention economy built their attentional baseline in an environment that rewarded short cycles and rapid information-grazing. AI tools then layer onto habits already formed, reinforcing a pattern rather than introducing one. Workers who entered professional life before the smartphone, by contrast, often built their concentration in environments that demanded it — long reading, uninterrupted problem-solving — and carry stronger baselines as a result.

The important and slightly ominous part is what happens over time. Those older baselines do not appear to be permanently protective. Workers in their forties and fifties who once prided themselves on hours of unbroken focus increasingly report that deep concentration is harder than it was a decade ago, and the more thoroughly AI threads into their daily work, the more their patterns drift toward those of younger colleagues. If fragmentation were mainly a generational trait, it would stay put with the generation. It doesn't. It behaves like a function of the environment, and as older workers accumulate exposure, the gap narrows. The direction of travel is convergence — not toward the older baseline, but toward a shared, fragmented one. For an organization, the lesson is that it cannot count on seasoned staff as a permanent reservoir of deep-focus capacity insulated by age. That reservoir depletes under the same conditions as everyone else's.

The Productivity Paradox

Here is the paradox at the center of it all: AI can raise measurable output while lowering the cognitive depth that produces the most valuable output — and standard productivity metrics are blind to the trade.

The visible half is easy to see. An engineer with an assistant closes more tickets and commits more code. A writer produces more words per hour. An analyst turns out more reports. Designers generate more concepts. By almost any quantitative measure an organization actually tracks, the AI-assisted worker looks more productive, and often genuinely is, for a large class of tasks.

The hidden half is that the nature of the output shifts. The engineer assembles and validates AI-generated components instead of reasoning an architecture through from first principles. The writer produces volume but wrestles more with the sustained logical development that separates strong long-form work from a pile of competent paragraphs. The analyst ships more reports while investigating the underlying questions less deeply. And depth is where the disproportionate value lives. The most consequential knowledge work is rarely high-volume production; it is insight — spotting the non-obvious pattern, catching the structural problem before it becomes a crisis, synthesizing distant ideas into something new. Those acts require exactly the condition AI erodes: holding a complex problem in mind long enough for the connections to surface.

This creates a measurement gap that sits at the heart of organizational analytics. Volume is trivial to count; depth is nearly impossible to. So dashboards fill with the metrics that are easy to capture — commits, tickets, words, reports — precisely the metrics AI inflates, while the capability that actually distinguishes valuable knowledge work goes uncounted. An organization can watch its productivity numbers climb while the thing those numbers were meant to proxy for quietly declines, and never see it in the data. The measurement doesn't just miss the problem; it actively disguises it as success.

The Long Shadow, and Why It's Hard to Escape

If chronic fragmentation were merely uncomfortable, it would be a wellness footnote. The concern is that it degrades capability. Research on attention associates chronic fragmentation with impaired acquisition of complex skills — because mastery is built through extended, effortful practice, and fragmented attention starves that practice of the sustained engagement it needs. It associates fragmentation with weaker creative problem-solving, since creative synthesis depends on holding disparate elements in mind together long enough for them to combine. And it associates fragmentation with heightened susceptibility to burnout, the exhaustion of never quite recovering from the last interruption while bracing for the next. These are losses of ability, not just comfort.

Which raises the reversibility question, and the honest evidence cuts both ways. Attention is trainable — the same plasticity that lets the environment fragment it should, in principle, let deliberate practice rebuild it, and there is reason to think capacity can be recovered. But the recovery appears to be slow and asymmetric. The consistent, if largely experiential, report is that fragmentation is fast to acquire and slow to shed: conditioning builds quickly under constant exposure and dissipates only reluctantly once the exposure stops. Whether that asymmetry reflects durable neural change or simply entrenched habit is exactly the sort of question we do not yet have the longitudinal evidence to answer.

This asymmetry is why individual countermeasures so often fail, and they fail through two reinforcing obstacles. The first is competitive pressure. A worker who switches off AI suggestions or silences notifications typically discovers that the resulting slowdown is real, and in a deadline-driven job the efficiency they've surrendered feels like a professional handicap. The tool comes back on. The second obstacle waits for those who persist: even after the external interruptions are gone, attention does not obediently return to depth. The brain, conditioned to expect the next prompt, treats the silence as something to be filled rather than a space to think in. Workers describe sitting with a hard problem in quiet as uncomfortable, even faintly anxious — the opposite of the productive absorption they remember. The two obstacles compound. The first keeps pulling the tools back before recovery can begin; the second ensures that even when they are removed, recovery is slow. Together they explain why willpower alone rarely works, and why the problem is better understood as structural than as a matter of personal discipline.

Who Is Responsible, and What Would Help

If the problem is structural, the response has to be too — and that raises questions of obligation that fall on four different actors.

Start with the tool designers, because they hold the most direct lever. The strong argument is that they should be obligated to optimize for attentional sustainability rather than raw engagement. The objection is that engagement is the business model. But the two may not be as opposed as they look. A tool that leaves its users cognitively depleted and burned out is not obviously winning in the long run, and a growing case holds that aligning product-success metrics with users' cognitive health is both commercially feasible — through designs that consolidate rather than scatter interruptions, that batch suggestions instead of streaming them, that make deep-work modes first-class features — and ethically required, given that the harm falls on the user while the benefit accrues to the vendor. At minimum, "engagement" is a poor proxy for "value delivered," and a firm that mistakes one for the other is optimizing the wrong number as surely as the organizations described above.

Employers hold the next lever, and their most consequential move is to fix what they measure. If volume metrics are the ones AI inflates, then leaning on them harder in an AI-saturated workplace guarantees the measurement gap widens. The alternative frameworks are not exotic: judging knowledge work by the quality and consequence of its best outputs rather than the count of its routine ones; protecting blocks of uninterrupted time as a genuine organizational asset rather than a perk; treating responsiveness as one value among several rather than the master metric. Companies experimenting with AI-free hours and notification blackout windows report modest gains in self-reported focus — modest, because the deeper obstacle is cultural. As long as availability reads as commitment and unavailability as risk, workers will quietly defect from their own protected time.

Regulators occupy the outermost and most contested ring. Their legitimate sphere is narrow but real: the same logic that produced rules around addictive design and children's screen exposure could extend to transparency about engagement optimization, particularly where tools are marketed as productivity aids while being tuned for retention. This is the least developed front, and reasonable people disagree about how far it should go.

And the individual worker? Not nothing — attention responds to deliberate practice, and personal boundaries help at the margin. But the recurring lesson of this chapter is that placing the whole burden on individual willpower, in an environment engineered against it, is a setup for failure and self-blame. Responsibility is shared, and it is heaviest where the leverage is greatest: with the people who design the tools and the organizations that decide what to measure.

What We Still Don't Know

It is worth being explicit about the edge of the map, because this is a domain where confident pronouncements outrun the evidence. The generative-AI era is only a few years old. We do not have longitudinal studies following the same workers across a decade of AI use. We cannot yet say whether the effects captured in short EEG experiments accumulate into durable change or wash out. We cannot cleanly separate AI's contribution from the smartphone-and-social-media substrate it arrived on. And we do not know whether what we are watching is a temporary adjustment — a workforce learning, as it once learned to manage email, to metabolize a powerful new tool — or a permanent downward shift in the human capacity for sustained attention.

Settling that would take the studies we don't yet have: long-term cohorts tracking attention as AI exposure accumulates; experimental designs that establish AI use precedes attentional change rather than following it; and neurophysiological work that measures not just the brain-on-tool state but the baseline that persists after the tool is set down. Until those exist, the intellectually honest position is that the in-the-moment costs of AI-assisted work are real and measurable, the environmental fragmentation of attention over the past two decades is real and measured, and the specific, durable, AI-caused rewiring of the human brain is a serious worry that remains, for now, unproven.

Summary

The famous numbers are a myth; the real decline is different. The "12 seconds to 8.25 seconds" attention-span figures trace to an unsourced statistic popularized around 2015 and debunked by the BBC — there is no instrument measuring a species-wide attention span year over year. The genuinely measured decline comes from Gloria Mark's research: average time on a single screen fell from ~2.5 minutes (2004) to ~47 seconds today. That measures a fragmenting environment, not a shrunken biological capacity.
AI's specific role is hard to isolate. Smartphones and social media fragmented attention for fifteen years before AI arrived at scale. Separating AI's contribution requires controlled comparisons, within-person tracking, and neurophysiological measurement — and must rule out reverse causation, since distractible people may simply adopt AI more. The environmental decline is well established; a durable, AI-specific brain change is not.
AI fragments through assistance, not just distraction. Four mechanisms — autocomplete suggestions, frictionless context switching, iterative prompting, and evaluation overhead — each consume cognitive resources, and unlike passive tools, AI initiates interaction, placing it on the active side of the interruption ledger.
Brain studies show real in-the-moment effects, not proven permanence. MIT's 2025 "Your Brain on ChatGPT" EEG study found LLM users had weaker neural connectivity, reduced alpha activity, and poorer recall of their own writing. This is a small, preliminary study showing a real difference while using the tool; the leap to permanent rewiring is not yet supported.
The interruption arithmetic is brutal. With ~23 minutes to recover full depth after a distraction (Mark) and interruptions arriving far more often, complete recovery never happens — deep focus becomes structurally, not just personally, unreachable.
The generational gap is real but closing. Younger workers report more fragmentation (conditioned earlier), but older workers' deep-focus baselines erode with AI exposure. Fragmentation behaves like a function of the environment, so the workforce converges toward a shared fragmented baseline rather than staying split by age.
The productivity paradox hides in the metrics. AI raises countable output (commits, words, reports) while eroding the depth that produces insight. Because volume is easy to measure and depth nearly impossible, dashboards inflate exactly the metrics AI boosts — disguising cognitive decline as success.
Recovery is slow and countermeasures fail structurally. Fragmentation builds fast and fades slowly. Individual fixes fail through two compounding obstacles: competitive pressure pulls the tools back on, and conditioned expectation prevents focus from returning even when interruptions are removed. The problem is structural, not a failure of willpower.
Responsibility is shared and heaviest where leverage is greatest. Tool designers should optimize for attentional sustainability over engagement; employers should measure the depth and consequence of work rather than its volume; regulators have a narrow role around transparency; and individuals help at the margin — but the burden should not rest on willpower alone.
The core question is still open. We lack the decade-long, well-controlled studies needed to know whether AI-saturated environments produce a temporary adjustment or a permanent shift in human attention. The honest verdict: the moment-to-moment costs are real, the environmental decline is measured, and the durable AI-caused rewiring remains a serious but unproven concern.

Sources

Two caveats on those sources, in the interest of faithful reporting: I wrote this offline from prior knowledge, so I could not confirm the exact URLs are live tonight — several (the BBC myth-busting piece, the HBR/Gloria Mark link, the MIT Media Lab page) are cited from memory and should be validated before publishing. The Gitnux/aggregator link is included deliberately as the subject of the myth critique, not as an authority. If you'd like, re-enable the web tools and I'll run a proper source-validation pass to lock the citations down.

Last updated: 2026-08-28

V2 (in progress) Previous: V1