The eight-second goldfish statistic has no source. Screen attention did fall from 150 to 47 seconds. What is measured, what is viral and what has changed.
The Statistic That Never Existed
How an eight-second number with no source became common knowledge
In May 2015 a marketing report from Microsoft Canada was reported around the world as proof that the human attention span had fallen to eight seconds, one second shorter than that of a goldfish [2]. ✓ Established The number was not a Microsoft finding. It appeared in the report as a footnote to a website that could not produce a source, and no study has ever measured it [1]. A decade later the figure still anchors conference keynotes, marketing decks and school policy debates. This report separates what has been measured from what has merely been repeated.
The document at the centre of the myth is a 54-page consumer-insights report titled “Attention Spans”, produced by Microsoft Canada's advertising research team in spring 2015. Its own methodology was modest: an online survey of 2,000 Canadians with game-like attention tasks, followed by electroencephalography of 112 volunteers watching media [2]. Its own conclusion was not that attention had collapsed. It found that heavy digital users showed intense bursts of attention at the start of a task that faded over time, while lighter users started slower and attended more to detail [2]. The eight-second figure, and the nine-second goldfish, appeared on an early page as context, with a footnote pointing to a website called Statistic Brain [1].
In 2017 the BBC World Service programme More or Less followed that footnote. Statistic Brain attributed its numbers to the US National Center for Biotechnology Information at the National Library of Medicine and to the Associated Press. Neither organisation could find any record of research supporting the figures, and Statistic Brain itself produced nothing when asked [1]. ✓ Established The BBC's conclusion was unambiguous: there was no scientific evidence that human attention spans were shrinking in the way claimed, and the goldfish figure had no research foundation either. Goldfish, for what it is worth, demonstrate associative learning that persists for months [1].
The number survived its debunking because it served several interests at once. For advertisers it justified shorter formats. For editors it was a perfect headline. For a public already primed by Nicholas Carr's 2008 Atlantic cover essay asking whether Google was making us stupid, which argued that online reading was eroding the capacity for contemplation, it confirmed a fear that was already in circulation [40]. By December 2024 that fear had a name: Oxford University Press chose “brain rot” as its Word of the Year after usage rose 230 percent between 2023 and 2024 and more than 37,000 people voted [25]. ✓ Established The dictionary defines it as the supposed deterioration of a person's mental state from overconsumption of trivial online content. The word “supposed” is doing real work in that definition.
The confusion that the myth exploits is a confusion between two different things. An attention span, in the clinical sense, is the capacity to sustain focus on a task over time. Attention switching is a behaviour: how often a person moves from one thing to another in a given environment. The first is a property of a brain. The second is a property of a brain in a setting. Gemma Briggs of the Open University told the BBC that “average attention span” is not a metric psychologists would try to measure, because how much attention a person applies depends on what the task demands [1]. A surgeon and a commuter scrolling a feed are not displaying different attention spans. They are in different rooms.
What has demonstrably accelerated is the environment. Philipp Lorenz-Spreen and colleagues showed in Nature Communications in 2019 that collective attention cycles have shortened across Twitter, Google Books, film ticket sales, Reddit and Wikipedia: a hashtag that stayed in the global daily top 50 for 17.5 hours on average in 2013 lasted 11.9 hours in 2016 [24]. ✓ Established Their model attributed the acceleration to rising content production competing for a finite pool of collective attention, not to any change in individual brains [24]. The same pattern appears in film. Cornell psychologist James Cutting, working from Barry Salt's database of more than 15,000 films, found that the average shot in English-language cinema fell from about 12 seconds in 1930 to about 2.5 seconds today [41]. Editors cut faster because they can. That is a fact about production, not about cognition.
Microsoft Canada's 2015 report surveyed 2,000 people and ran EEG on 112; it did not measure an eight-second span and did not claim to [2]. The figure came from Statistic Brain, which cited the US National Library of Medicine and the Associated Press. Neither had any record of the research, and Statistic Brain produced none when the BBC asked in 2017 [1]. Every subsequent repetition of the number, in marketing, journalism and policy, traces back to that empty footnote.
What Has Actually Been Measured
Screen attention has collapsed; measured attention capacity has not
Gloria Mark of the University of California, Irvine, has logged how long people stay on a single screen since 2004. The average was about two and a half minutes then, about 75 seconds around 2012 and 47 seconds in studies run between 2016 and 2020 [3]. ✓ Established Over the same period, standardised tests of attention capacity in adults have shown no decline and, in the largest meta-analysis available, a modest improvement [9][10]. ◈ Strong Evidence Both findings are robust. They are measuring different things.
Mark's method matters because it is objective. Her early studies followed office workers with stopwatches, recording every shift of attention from one window, document or device to another. Later work used logging software running unobtrusively in the background as people went about their normal days, in workplaces, among students and at home [3]. The 47-second figure is the average of her last published studies, which ranged from 44 to 50 seconds, and it has been replicated by other research groups using the same logging techniques [3]. ✓ Established The trajectory is monotonic: roughly 150 seconds in 2004, 75 seconds by 2012, under a minute since 2016 [3]. Few behavioural measures in the social sciences have moved so far, so consistently, in so short a time.
It is important to be precise about what the measure captures. It records switching behaviour on screens: how long a person stays on one thing before moving to another. Mark's own data show that close to half of those switches are self-initiated, with no external notification or interruption involved [3]. The measure does not test whether a person can sustain attention when the task requires it. It tests how long they do so when nothing requires it and everything else is one tap away. Mark herself draws this distinction. Her finding is about the experience of attention in a digital environment, not a claim that the underlying capacity has degraded [3].
The capacity question has its own literature, and it points the other way. Edward Vogel, professor of psychology and neuroscience at the University of Chicago, told the Wall Street Journal in 2017: “I've been measuring college students for the past 20 years. It's been remarkably stable across decades” [10]. Michael Posner, who mapped the brain networks underlying attention, said there was no real evidence that attention had changed since it was first measured in the late nineteenth century, and the neurologist Marcus Raichle agreed [10]. ◈ Strong Evidence In 2015 Francesca Fortenbaugh and colleagues tested 10,430 people aged 10 to 70 on a continuous performance task and found that sustained-attention ability rises through the twenties and thirties, peaks in the early forties and declines thereafter, with strategy shifting toward caution with age [11]. The developmental curve is clear. A secular collapse is not in it.
Denise Andrzejewski, Elisabeth Zeilinger and Jakob Pietschnig at the University of Vienna pooled every published administration of the d2 Test of Attention they could find across three decades and asked whether there was a Flynn effect for attention [9]. There was, for adults: moderate generational gains in concentration performance. Children showed no such gain. Their overall error rate rose while their processing speed increased, a pattern consistent with trading accuracy for speed rather than with a loss of capacity [9].
The Vienna meta-analysis is the closest thing the field has to a direct test of the shrinking-span hypothesis, and it refutes the strong version. If adults had lost the ability to concentrate, three decades of standardised testing in 32 countries would show it. Instead the trend runs the other way, in line with the broader Flynn effect in cognitive test scores [9]. ✓ Established The one qualified signal is in children, where more errors alongside faster responding suggests a changed speed-accuracy strategy. That is a real finding and it deserves follow-up. It is not the same as the claim that a generation can no longer pay attention, and the authors do not make that claim.
Every honest statistic in this debate falls into one of two boxes. Capacity measures, such as the d2 test, continuous performance tasks and laboratory vigilance paradigms, ask what a person can do under instruction, and in adults they have been flat or rising for thirty years. Behaviour measures, such as Mark's screen logs, PISA's distraction questions and platform time-on-app data, ask what people actually do in a designed environment, and they have moved dramatically. Policy that targets capacity tells people to train their attention. Policy that targets behaviour changes the environment. Confusing the two produces advice for individuals where the evidence points at design.
The capacity literature has limits that its defenders should acknowledge. Standardised attention tests are brief laboratory tasks, typically ten to twenty minutes long, and they may not capture the hour-long sustained reading that a novel or a proof requires. The d2 test measures selective attention and concentration under time pressure, not the ability to stay with a single difficult text. Many of the stable samples Vogel describes are university students, a selected population. Fortenbaugh's 10,430 participants were self-selected volunteers on a website [11]. None of this reverses the findings. It does mean that “capacity is stable” is a statement about what the available instruments measure, and the instruments were not designed to detect a decline in long-form reading.
Reading is where behavioural indicators look worst. In 2024 the average US twelfth-grade reading score on the National Assessment of Educational Progress fell to 283 on a 500-point scale, the lowest in more than three decades; 35 percent of seniors read at or above the proficient level and 32 percent fell below basic [28]. ✓ Established Among adults, 48.5 percent read at least one book for pleasure in 2022, down from 52.7 percent in 2017 and 54.6 percent in 2012 [29]. These are measures of what people do, not of what they can do, and they are consistent with displacement: hours that once went to books now go elsewhere. They are not evidence that the capacity to read has been lost. They are evidence that it is being exercised less, which over time may amount to the same thing for the individuals concerned.
The Real Cost of Switching
The measurable harm is in the interruptions, not the span
If attention capacity is intact, the damage done by a fragmenting environment must show up somewhere else. It does: in the cost of each switch. In 2001 Joshua Rubinstein, David Meyer and Jeffrey Evans showed that moving between tasks imposes a time penalty through two mental stages, goal shifting and rule activation, and that the penalty grows with task complexity [6]. ✓ Established Two decades of workplace observation since then have converted that laboratory finding into a bill.
The mechanism is well understood. When a person switches tasks, the brain's executive control system must first decide to do this rather than that, then deactivate the rules of the old task and load the rules of the new one. Each stage takes time, and both take longer when the tasks are unfamiliar or complex [6]. Rubinstein, Meyer and Evans estimated that a worker unable to concentrate for even ten minutes at a stretch could lose as much as 20 to 40 percent of potential efficiency to switching costs [6]. The 40 percent figure is an upper bound for complex tasks, not a universal average, and it has been quoted far beyond what the experiment supports. The direction of the finding, however, has held up in every replication: switching is never free.
Mark's 2005 field study put the cost in minutes. Her team shadowed 24 information workers at an outsourcing firm, seven managers, nine analysts and eight developers, for about three and a half days each, roughly 700 hours of observation [4]. ✓ Established Fifty-seven percent of the workers' working spheres were interrupted, and once interrupted, a task took on average more than 25 minutes to resume, with more than two other activities intervening in between [4]. The widely quoted “23 minutes and 15 seconds” is a corruption of this result. The peer-reviewed figure is longer. Three years later Mark, Daniela Gudith and Ulrich Klocke tested interruption experimentally and found something subtler: interrupted participants finished tasks faster with no loss of quality, but reported significantly more stress, frustration, time pressure and effort [5]. ✓ Established People compensate for interruption by speeding up. They pay for the speed in cortisol.
Sophie Leroy added the missing piece in 2009. In a series of field and laboratory studies published in Organizational Behavior and Human Decision Processes, she showed that when people leave a task unfinished, part of their attention stays with it, and their performance on the next task suffers as a result [7]. ◈ Strong Evidence She called the phenomenon attention residue, and found it was worst when the unfinished task carried anticipated time pressure. A brief glance at something unrelated was enough to degrade performance on a demanding puzzle when participants returned to it [7]. The practical implication is that the cost of a switch is not paid at the moment of switching. It is paid for minutes afterwards, in a tax on whatever comes next.
We spend an average of just 47 seconds on any screen before shifting our attention.
— Gloria Mark, Chancellor's Professor of Informatics, University of California, Irvine, Attention Span, 2023Whether heavy switchers become worse at filtering is a separate question, and the evidence is weaker than its fame. In 2009 Eyal Ophir, Clifford Nass and Anthony Wagner at Stanford surveyed 262 students on their media habits and tested the 19 heaviest and 22 lightest media multitaskers [8]. The heavy group was worse at ignoring irrelevant information and slower to switch between tasks: as Wagner put it, they were unable to filter out what was not relevant to their current goal [8]. ◈ Strong Evidence The study was cross-sectional and small, and it cannot say whether multitasking caused the deficit or whether people with weaker filtering gravitate to multitasking. Later attempts to reproduce the pattern have been mixed. It remains one of the most cited findings in the field and one of the least settled.
The replication record elsewhere is a caution. In 2017 Adrian Ward and colleagues reported that the mere presence of a participant's own smartphone, face down and silent, reduced available working memory and fluid intelligence, most sharply among the most dependent users [16]. The “brain drain” result was reported everywhere. Two subsequent direct replications, in 2021 and by Ruiz Pardo and Minda in 2022, found no difference between phone-location conditions on either task [16]. ⚖ Contested The original study may have been right about some populations and wrong about others, or it may have been a false positive. Either way, the most widely shared single finding about phones and cognition does not currently replicate, and anyone citing it as settled is behind the literature.
The scale of exposure is what turns small per-switch costs into a population-level phenomenon. Internet users worldwide spent an average of 141 minutes per day on social media in February 2025, down slightly from 143 minutes a year earlier and up from about 90 minutes in 2012 [36]. ✓ Established Among US teenagers, 46 percent told Pew Research Center in late 2024 that they are online almost constantly; 73 percent use YouTube every day and about six in ten use TikTok daily, with 16 percent describing their TikTok use as almost constant [26]. ✓ Established At 47 seconds per screen, two hours and twenty minutes of daily social media use implies roughly 180 attention shifts a day on those platforms alone, before counting work, messaging and the rest of the screen day.
Children, Screens and the Developing Brain
The effects are real, small and contested, and the samples are not what the headlines imply
The strongest version of the attention panic concerns children: that screens are rewiring developing brains and producing a generation unable to focus. The evidence is more specific than that. In the largest paediatric neuroimaging cohort in the world, more screen time at ages nine and ten predicted slightly higher ADHD symptoms two years later [13]. ◈ Strong Evidence In the same cohort, screen time showed no relationship with functional brain development over the same two years [14]. ⚖ Contested Both results are published, both are large, and they do not say what either camp wants them to say.
The Adolescent Brain Cognitive Development study follows 11,878 American children with repeated MRI scans, behavioural assessments and parent reports. In October 2025 Qiulu Shou, Masatoshi Yamashita and Yoshifumi Mizuno of the University of Fukui published an analysis of 10,116 of those children at ages nine to ten and 7,880 at two-year follow-up in Translational Psychiatry [13]. Children who spent more time on screens had slightly more severe ADHD symptoms at baseline, and longer screen time at baseline predicted higher symptom scores two years later after adjusting for the starting level [13]. ◈ Strong Evidence Greater screen exposure was also associated with lower total cortical volume, lower volume in the right putamen and thinning in several frontal and temporal regions, and cortical volume partially mediated the link between screen time and symptoms [13]. The authors describe the magnitude of every one of these associations as small.
Two years earlier, Patrick Miller, Kathryn Mills, Matti Vuorre, Amy Orben and Andrew Przybylski had published a different analysis of nearly 12,000 children from the same cohort in Cortex [14]. They asked whether self-reported screen engagement related to functional brain connectivity measured by MRI, and whether it affected neural maturation over two years. It did not. The results did not support the prediction that functional brain organisation is related to digital screen engagement, and exploratory analyses showed no significant effect on neural maturation over the period [14]. ◈ Strong Evidence The two papers are not contradictory. One measured structure and symptoms and found small associations; the other measured functional connectivity and found none. Together they describe an effect that is detectable in some measures, absent in others and modest everywhere.
The question of how modest is best answered by the largest analysis of adolescent technology use ever conducted. In 2019 Amy Orben and Andrew Przybylski applied specification curve analysis, which runs every defensible version of a statistical model rather than one chosen version, to three datasets covering 355,358 adolescents [15]. Digital technology use explained at most 0.4 percent of the variation in well-being [15]. ✓ Established Within the same data, eating potatoes showed a comparable negative association, wearing glasses a larger one, and being bullied or smoking cannabis associations 4.3 and 2.7 times more negative [15]. The authors concluded that the effects were too small to warrant policy change. Whether or not one accepts that conclusion, the number itself has not been overturned.
The September 2025 Psychological Bulletin meta-analysis by Nguyen and colleagues found that greater use of TikTok, Reels and comparable feeds was associated with poorer overall cognition (r = −0.34), with attention (r = −0.38) and inhibitory control (r = −0.41) showing the strongest associations and reasoning showing none [12]. The attention figure has been widely reported as proof that short-form video damages focus. The cognitive sub-analysis drew on 14 studies, every one conducted in China, almost all cross-sectional [12]. A correlation of that size is real and worth attention. It cannot say which way the arrow points.
A correlation of −0.38 is, by the standards of behavioural science, moderate. It is also compatible with at least three causal stories. Heavy short-form video use may erode attentional control. People with weaker attentional control, including the roughly one in nine American children who have ever received an ADHD diagnosis, may be drawn to feeds that deliver a new stimulus every few seconds. Or a third factor, such as sleep loss or family circumstance, may drive both. Cross-sectional data cannot separate these, and the meta-analysis's authors say so. The geographic concentration adds a further caveat: Chinese university samples using Douyin are not obviously representative of adolescents in Osaka, Lyon or São Paulo. The finding is a reason to run longitudinal and experimental studies. It is not a finding that those studies have already been run.
The neuroscience is thinner still. A 2024 Zhejiang University study fitted 48 participants with EEG caps during the Attention Network Test and found that heavy short-form video users showed poorer executive attention and reduced theta activity in the prefrontal cortex [17]. The result is biologically plausible and the sample is tiny. Meanwhile the clinical signal that most alarms parents, rising ADHD diagnosis, has its own explanations. The CDC reported in October 2024 that 15.5 million US adults, 6.0 percent, had a current ADHD diagnosis in 2023, about half of them diagnosed at 18 or older, and that 71.5 percent of those on stimulants had trouble filling prescriptions during shortages [27]. ✓ Established Adult diagnosis at that scale reflects widened criteria, telehealth access and recognition of a condition long underdiagnosed in women. It cannot be read as a measure of what screens have done to brains.
The honest summary of the paediatric evidence is that effects are small, consistent in direction across the better studies and silent on causation. Small is not the same as negligible: an effect of 0.4 percent of variance, applied to a cohort of tens of millions of children and several hours a day, can move the tail of a distribution, which is where clinical ADHD lives. But small effects also cut the other way. If the arrow runs from vulnerable children toward heavy use rather than from heavy use toward vulnerability, then restricting screens will leave the vulnerability untouched. The research design that could settle this, randomised reduction of use in children followed over years, has barely begun. Policy is being made ahead of it.
The one secular finding in children that does not depend on self-reported screen time comes from the Vienna meta-analysis. Across 32 countries and three decades of d2 testing, children's overall error rates rose while adults' concentration scores improved [9]. ◈ Strong Evidence This is a change in measured performance under standardised conditions, not a correlation with a questionnaire. It is consistent with the ABCD finding of slightly elevated attention symptoms, and it is consistent with the Karolinska Institute's 2023 position that digital tools in classrooms impair rather than enhance learning [34]. It is also consistent with children simply responding faster and more carelessly on a timed test than their predecessors did. The data do not distinguish between those readings. What they rule out is the claim that nothing at all has changed in how children perform attention tasks.
The Classroom Evidence
Nine countries, one directly measured problem
The 2022 round of the OECD's PISA assessment asked 15-year-olds in 81 systems about digital distraction. Across OECD countries 65 percent reported being distracted by their own devices in at least some mathematics lessons and 30 percent in most or every lesson; students distracted in most lessons scored 15 points lower in mathematics than those who rarely were [18]. ✓ Established Unlike the goldfish, this is a measurement, taken from hundreds of thousands of students, and it is the evidence base on which most of the world's phone policy now rests.
The PISA figures describe a problem with an unusually clear geography. Fifty-nine percent of students across the OECD said their attention was diverted by other students' phones, tablets or laptops in at least some mathematics lessons, and 45 percent said they feel anxious when their phone is not near them [18]. ✓ Established The proportion reporting own-device distraction exceeded 80 percent in Argentina, Brazil, Canada, Chile, Finland, Latvia, New Zealand and Uruguay [18]. In Japan only 18 percent and in Korea 32 percent reported being distracted by classmates' devices, among the lowest figures recorded [18]. Japan and Korea also sit at the top of the mathematics rankings. The correlation does not prove that quiet classrooms cause high scores, but it places the two highest-performing large systems at the low-distraction end of the distribution.
UNESCO's 2023 Global Education Monitoring Report concluded that almost one country in four had already banned smartphones in schools, and that having a phone nearby with notifications switched on was enough to pull attention from the task at hand [19]. ✓ Established It recommended bans where technology did not clearly support learning. The recommendation landed: UNESCO's tracking shows 60 education systems, 30 percent of those monitored, had bans in place by the end of 2023 and 79 systems, 40 percent, by the end of 2024 [19]. Few education policies have spread as fast. The question is whether they work, and here, unusually for this debate, there is quasi-experimental and registry evidence.
The foundational study is Louis-Philippe Beland and Richard Murphy's analysis of 91 schools in four English cities, published by the LSE's Centre for Economic Performance in 2015 and in Labour Economics in 2016. After schools introduced phone bans, test scores of 16-year-olds rose by 6.4 percent of a standard deviation, the equivalent of about five additional school days a year [20]. ✓ Established The gain was driven entirely by low-achieving pupils; high achievers showed no significant change [20]. Sara Abrahamsson's 2024 study of more than 400 Norwegian middle schools, using national health and education registers, found that bans reduced girls' consultations for psychological symptoms, lowered bullying for both sexes and raised girls' GPA and externally graded mathematics results by about 0.22 standard deviations, with the largest gains among girls from low-income families [21]. ◈ Strong Evidence Two countries, two methods, one pattern: the students who were being distracted most benefited most.
The Dutch evidence is administrative rather than experimental but points the same way. The Netherlands issued national guidance in January 2024 that phones should be kept out of classrooms, and almost all schools complied. An evaluation by the Kohnstamm Instituut for the Dutch government, published in July 2025, surveyed 317 secondary schools: three-quarters reported improved concentration, nearly two-thirds a better social climate and one-third better academic results, while the effect in primary schools was minimal [23]. ◈ Strong Evidence The English SMART Schools study is the counterweight. Victoria Goodyear's team at the University of Birmingham compared 1,227 pupils aged 12 to 15 in 20 schools with restrictive policies and 10 with permissive ones, publishing in The Lancet Regional Health – Europe in February 2025. They found no significant differences in mental wellbeing, sleep, physical activity, classroom behaviour or attainment between the two policy types [22]. ✓ Established Total daily phone and social media use, on the other hand, predicted worse outcomes on every one of those measures [22].
The SMART result is often read as showing that bans fail. It shows something more precise: a school ban covers six or seven hours of a day in which adolescents use phones for four to six, and the harms the study measured, poor sleep and low wellbeing, accrue mostly in the hours the ban does not touch. Goodyear's own conclusion was that school bans alone are not enough [22]. Set beside Beland and Murphy and Abrahamsson, the picture is coherent. Bans improve what happens in the classroom, especially for the pupils most prone to distraction, and leave untouched what happens after the bell. That is exactly what one would predict if the problem is the environment rather than the child.
The country comparison exposes how different the policy instruments are. France has prohibited phones in primary and lower-secondary schools since 2018 under Article L. 511-5 of the Education Code; the “portable en pause” scheme, which stores phones for the whole school day, was trialled with more than 32,000 pupils in 2024–25, generalised to every collège in September 2025 with reported gains in concentration, school climate and cyberbullying, and President Macron announced on 28 November 2025 that lycées would follow from September 2026 [31]. ✓ Established South Korea amended its Elementary and Secondary Education Act on 27 August 2025 by 115 votes to 31, banning classroom use from 1 March 2026 with exceptions for disability and teaching, and with no penalties [32]. Brazil's law of 13 January 2025 covers classrooms and breaks nationwide; 64 percent of Brazilian schools already had some restriction and 28 percent an outright ban by 2023 [37]. In the United States, 32 states and the District of Columbia now require districts to ban or restrict phones, two years after Florida passed the first such law [30].
Three countries have gone beyond the school gate. Sweden, whose fourth graders' PIRLS reading scores fell from 555 in 2016 to 544 in 2021, reversed its digital-first strategy after the Karolinska Institute declared that there was clear scientific evidence that digital tools impair rather than enhance student learning; Education Minister Lotta Edholm ended mandatory devices in preschools and the government budgeted 685 million kronor for textbooks in 2023 with 500 million a year to follow [34]. ✓ Established China's Cyberspace Administration drafted “minor mode” rules in August 2023 capping daily device use at 40 minutes under age eight, one hour from eight to sixteen and two hours from sixteen to eighteen, with no mobile use between 10 pm and 6 am [38]. And in Japan, where the education ministry has treated phones as prohibited in principle since 2009, the city of Toyoake in Aichi passed the country's first all-resident ordinance by 12 votes to 7, in force from 1 October 2025: a two-hour daily guideline for leisure screen use, device cut-offs at 9 pm for primary pupils and 10 pm for older students, and no penalties [33]. ✓ Established Mayor Masafumi Kouki described the limit as “a guideline, not a blanket rule” [33]. Australia, from 10 December 2025, became the first democracy to require platforms to remove under-16 accounts, on pain of fines up to A$49.5 million [39]. Nine jurisdictions, nine instruments, one measured problem.
The Policy Response
What the evidence supports, what it does not, and what nobody has tested
Governments are converging on restriction faster than the research can evaluate it. Of the instruments now in use, school-day phone bans have the best evidence behind them, with gains concentrated among the students most at risk [20][21][23]. ◈ Strong Evidence Age-based social media bans and statutory screen-time caps are untested at scale. And the one intervention that targets the mechanism the research actually identifies, the design of the switching environment itself, is barely on the agenda.
The policy menu now has six items. School-day bans, in force in France, the Netherlands, Brazil, South Korea and most US states, remove devices from the six or seven hours in which distraction is directly measured [31][23][37][32][30]. Age-gating, pioneered by Australia, removes a category of platform from a category of user [39]. Statutory time caps, enforceable in China's minor mode and advisory in Toyoake, limit total exposure [38][33]. Curriculum reversal, as in Sweden, reduces the role of screens in the learning itself [34]. Design regulation would alter the features that produce switching: autoplay, infinite scroll, notification cadence. And individual-level advice, from digital detoxes to focus apps, asks the user to resist the environment. These are not variations on one policy. They target different parts of the causal chain, and the evidence for each is very different.
For school bans the evidence is strongest and most specific. Three independent designs, Beland and Murphy's difference-in-differences across 91 English schools, Abrahamsson's registry analysis of 400 Norwegian schools and the Dutch 317-school evaluation, agree on direction and on who benefits: low achievers, girls and pupils from poorer families [20][21][23]. ✓ Established The SMART Schools null result on wellbeing does not contradict this. It bounds it. Bans improve attainment and classroom climate during school hours and do not, on their own, improve sleep or mental health, because those outcomes are determined by use outside school [22]. A policymaker who promises that a school ban will fix adolescent wellbeing is overclaiming. One who promises better concentration and higher scores for struggling pupils is on solid ground.
Age-based bans and time caps are a different matter. Australia's under-16 ban took effect on 10 December 2025 with monthly compliance reporting to the eSafety Commissioner, and the earliest outcome data will not exist for years [39]. China's minor mode caps are enforceable in principle but parents may override them, and no independent evaluation has been published [38]. Toyoake's ordinance carries no penalties and its mayor calls it a guideline [33]. These are experiments in progress, and the honest assessment is that they are plausible, popular and unproven. The risks are also real: circumvention through false ages and borrowed devices, exclusion of adolescents who depend on online communities, and the displacement of use into less visible channels.
| Policy Instrument | Evidence Base | Assessment |
|---|---|---|
| School-day phone bans | Quasi-experimental and registry evidence from England, Norway and the Netherlands shows higher scores and better concentration, concentrated among low achievers and girls [20][21][23]. No measured effect on wellbeing or sleep, which are driven by use outside school [22]. | |
| Under-16 social media bans (Australia model) | Targets the platforms with the shortest switching cycles, but took effect in December 2025 and has no outcome data [39]. Enforcement through age assurance is untested and exclusion risks for vulnerable adolescents are documented. | |
| Statutory screen-time caps (China, Toyoake) | China's minor mode sets enforceable age-graded limits with parental override and no published evaluation [38]; Toyoake's two-hour guideline carries no penalties [33]. Plausible mechanism, no evidence of effect yet. | |
| Design regulation of switching features | The only instrument aimed at the mechanism the switching research identifies: autoplay, infinite feeds and notification cadence [3][6]. Strong mechanistic evidence, almost no field trials, and the weakest political traction of any option. | |
| Individual advice and focus training | Leroy's ready-to-resume plans and batching strategies have laboratory support [7], but they place the burden on the user against an environment designed to defeat them. Necessary; not sufficient; no population-level effect demonstrated. |
Sweden's curriculum reversal is the most distinctive response because it targets the learning itself rather than the device. The country had made digital tools mandatory in preschools and relied heavily on tablets through primary school. After PIRLS recorded an 11-point fall in fourth-grade reading between 2016 and 2021, and after the Karolinska Institute's statement that the focus should return to printed textbooks and teacher expertise rather than unvetted digital sources, the government reversed the mandate, ended digital learning for children under six and funded textbooks with 685 million kronor in 2023 [34]. ✓ Established The evidence that this will raise reading scores is indirect. The evidence that the previous policy had coincided with their fall is direct. For a government, that asymmetry is usually enough.
What is striking across all nine jurisdictions is how little of the policy addresses the mechanism the research describes. Mark's 47 seconds, Rubinstein's switching costs, Leroy's attention residue and Lorenz-Spreen's accelerating collective attention all point at the same thing: an environment that offers a new stimulus every few seconds and makes switching frictionless [3][6][7][24]. ◈ Strong Evidence None of the bans, caps or curriculum changes alters that environment for the hours in which people are inside it. They remove people from the room for part of the day. The design of the room is left to the companies that built it, and those companies, as earlier OsakaWire reporting on the attention economy has documented, are paid by the switch.
A policy built on the goldfish treats the problem as a deficit in people and prescribes effort: focus training, detoxes, willpower. A policy built on the switching research treats the problem as a property of the environment and prescribes design: fewer stopping-cue removals, slower feeds, batched notifications, defaults that favour completion over interruption. The school bans that are spreading worldwide sit between the two. They change the environment, but only by removing it for six hours, and the SMART Schools result shows what that leaves untouched. The instrument the evidence most clearly supports, regulating the features that produce switching, is the one no government has yet seriously attempted.
The Debate
Degradation or moral panic: where the two camps actually disagree
Two serious positions now contend over attention. One holds that a measurable degradation is under way, visible in screen logs, short-form video correlations, paediatric cohorts and reading scores. The other holds that the panic has outrun the data, that capacity measures are stable, effect sizes tiny and the causal arrow unproven [9][35]. ⚖ Contested Both camps reject the goldfish. The dispute is about what the real numbers mean.
The degradation case assembles the behavioural record. Time on a single screen has fallen by two-thirds since 2004 [3]. Heavier short-form video use correlates with worse attention and inhibitory control across nearly 100,000 participants [12]. More screen time at age nine predicts more ADHD symptoms at eleven, with structural brain correlates [13]. Children's error rates on a standardised attention test have risen over three decades [9]. US twelfth-grade reading scores are at a 30-year low and adult pleasure reading is down six points in a decade [28][29]. ◈ Strong Evidence Sweden's leading medical university has stated that digital tools impair learning [34]. Each of these is a real finding. Taken together, the degradation camp argues, they describe a trend too consistent to be coincidence.
The sceptical case assembles the capacity record and the effect sizes. Adults' attention test scores have risen, not fallen, since 1990 [9]. Researchers who have measured college students for decades report stability [10]. The largest functional imaging analysis of children found no relationship between screen time and brain development [14]. Technology use explains 0.4 percent of adolescent well-being, less than wearing glasses [15]. The headline attention correlation rests on 14 Chinese studies [12]. The most famous phone-and-cognition experiment failed to replicate twice [16]. ◈ Strong Evidence Candice Odgers, reviewing Jonathan Haidt's The Anxious Generation in Nature in 2024, wrote that hundreds of researchers had searched for large effects and found “a mix of no, small and mixed associations”, and that correlation was being mistaken for causation [35]. Each of these is also a real finding.
The Case That Attention Is Degrading
Mark's objective logging shows 150 seconds per screen in 2004, 75 in 2012 and 47 since 2016, replicated by independent groups [3].
Across 98,299 participants, heavier use is associated with worse attention (r = −0.38) and inhibitory control (r = −0.41) [12].
In 7,880 ABCD children, more screen time at nine predicted higher ADHD symptoms at eleven, with lower cortical volume as a partial mediator [13].
Three decades of d2 testing across 32 countries show rising overall errors in children while adults improved [9].
The Case That the Panic Outruns the Data
The same d2 meta-analysis shows moderate generational gains in adult concentration across 21,291 people [9]; Vogel, Posner and Raichle report decades of stability [10].
Technology explains at most 0.4 percent of adolescent well-being variance, comparable to eating potatoes; bullying is 4.3 times more harmful [15].
Nearly 12,000 ABCD children showed no relationship between screen engagement and functional connectivity or neural maturation [14].
The attention estimate rests on 14 cross-sectional Chinese studies [12]; the “brain drain” experiment failed two direct replications [16].
Carr's 2008 essay followed identical anxieties about television, radio, novels and the telegraph, none of which produced measurable cognitive decline [40].
The two camps agree on more than their rhetoric suggests. Nobody serious defends the eight-second figure. Nobody disputes that switching has costs, that interruption raises stress or that classroom distraction is real and measured [6][5][18]. ✓ Established Both accept that school bans raise low achievers' scores and that they do not, alone, improve wellbeing [20][22]. The shared ground is, in fact, most of the evidence base. What remains contested is a narrower set of questions: whether the behavioural changes reflect or produce changes in capacity, whether small effects in children warrant precautionary policy and whether correlational findings in adolescents can be read causally.
The disagreement over causation is the deepest. Haidt argues that converging correlational, longitudinal and experimental evidence, together with the timing of the post-2012 shift in adolescent outcomes, is sufficient to infer cause. Odgers replies that the same pattern is consistent with economic hardship after the 2008 crisis and with increased reporting, and that effect sizes this small have never in the history of epidemiology been accepted as proof of a population-level epidemic [35]. ⚖ Contested On attention specifically, the strongest prospective evidence, the ABCD finding, is small and partly mediated by brain structure that may itself reflect pre-existing differences [13]. The randomised trials that would settle the question, reducing children's screen use for years and measuring attention afterwards, have not been done at scale. Until they are, the causal claim is a hypothesis with supporting correlations, not a finding.
I've been measuring college students for the past 20 years. It's been remarkably stable across decades.
— Edward Vogel, Professor of Psychology and Neuroscience, University of Chicago, Wall Street Journal, February 2017Measurement is the other fault line. Almost every study in the children's literature relies on self-reported screen time, which correlates weakly with logged use. The ABCD analyses, the Orben and Przybylski specification curve and most of the 71 short-form video studies all inherit that weakness [13][15][12]. The capacity literature has the opposite problem: precise measurement of a narrow construct. The d2 test takes a few minutes and measures selective attention under time pressure [9]. It cannot detect whether people can still read a 400-page book. When one side cites behavioural data of doubtful precision and the other cites precise data on a construct that may not be the one in dispute, each can be right about its own numbers and wrong about the argument.
The 2025 meta-analysis reports r = −0.38 between short-form video use and attention, a moderate association by behavioural standards, drawn from 14 cross-sectional studies conducted in China [12]. The Vienna meta-analysis reports rising adult d2 scores from 1990 to 2021 across 32 countries [9], and the Oxford Internet Institute's Cortex analysis of nearly 12,000 children found no link between screen engagement and brain development [14]. The correlation is real. The causal direction, the generalisability and the relationship to measured capacity are all unresolved.
What the Evidence Tells Us
Three different things are called attention, and only two of them have moved
The attention debate is three debates conducted in one vocabulary. Attention capacity, measured by standardised tests, has been stable or rising in adults for thirty years [9][10]. ✓ Established Attention behaviour, measured by screen logs and classroom surveys, has changed beyond recognition [3][18]. ✓ Established And the attention span of popular discourse, eight seconds and a goldfish, was never measured at all [1]. ✓ Established Policy that keeps these apart has a real evidence base. Policy that conflates them aims at the wrong target.
What is established beyond reasonable dispute can be stated in a few sentences. The eight-second figure has no source [1]. Time on a single screen before switching has fallen from about 150 seconds to 47 since 2004 [3]. Each switch costs time and each interruption costs stress; resuming interrupted work takes more than 25 minutes on average and unfinished tasks leave a residue that degrades the next one [6][4][5][7]. Two-thirds of 15-year-olds in the OECD are distracted by their own devices in class, those distracted most score 15 points lower, and removing phones raises the scores of the pupils who were struggling [18][20][21]. Adults' performance on standardised attention tests has not declined [9][10]. ✓ Established These findings come from logging, registry data, international assessment and meta-analysis. None of them depends on a footnote.
What is strongly supported but not settled concerns children. Heavier screen use at nine predicts slightly more attention symptoms at eleven, with small structural correlates [13]. Heavier short-form video use correlates moderately with poorer attention in Chinese samples [12]. Children's error rates on attention tests have risen over thirty years while adults' performance improved [9]. ◈ Strong Evidence Each of these is consistent with harm. Each is also consistent with reverse causation, with changed test-taking strategy or with confounders that no cross-sectional design can exclude. The honest position is that the signal is there, it is small, and the studies that could confirm it as causal have not been run.
What is contested is the interpretation that ties these findings into a single story of cognitive decline. The degradation camp reads falling screen attention, rising correlations and falling reading scores as one phenomenon. The sceptical camp reads stable capacity measures, tiny effect sizes and failed replications as proof that the phenomenon is a panic [35][16]. ⚖ Contested The evidence reviewed here supports neither reading in full. It supports a third: that capacity has not changed, that the environment has changed enormously, and that the costs of operating a stable brain in an environment engineered for switching are real, measurable and currently borne by individuals.
Attention has not shrunk. The number of demands on it has multiplied, and the friction between them has been engineered away. A brain that switches every 47 seconds in 2020 and every 150 seconds in 2004 is not a smaller brain; it is the same brain in a room with more doors. The research on switching costs says each door has a toll. The research on classroom bans says closing the doors for six hours helps the people who were walking through them most. The research on capacity says the brain itself is fine. Put together, those three findings describe a design problem, and design problems have design solutions. The goldfish framing turns a design problem into a character flaw, which is precisely why it has been so useful to the people who build the rooms.
The policy implications follow directly. Interventions that remove the environment for part of the day, as school bans do, have measured benefits for the students most affected and should be judged on those benefits rather than on wellbeing outcomes they were never positioned to deliver [20][21][22]. Interventions that cap total exposure, as in China and Toyoake, or remove platforms from age groups, as in Australia, target plausible mechanisms and should be evaluated rather than assumed to work [38][33][39]. Interventions that redesign the switching environment itself, the only ones aimed at the mechanism the research identifies, have the strongest mechanistic case and the least political momentum [3][6]. ◈ Strong Evidence And interventions that ask individuals to try harder are supported by the laboratory and undermined by the arithmetic: 180 switches a day on social media alone is not a willpower problem.
The goldfish, it turns out, remembers things for months. Humans, on the available evidence, can still concentrate as well as they ever could, when the room allows it. The honest statistic in this debate is not eight seconds. It is 47, and it describes the screens, not the brains. Ten years after a marketing footnote became common knowledge, the research record is clear enough to say what has changed and what has not. Attention capacity has held. The environment it operates in has been rebuilt around interruption, and the costs of that rebuilding, in minutes, in stress and in the test scores of the students least able to resist it, are measured, documented and growing. That is the finding. It does not need a fish.