Nine of eleven OECD countries recorded annual declines in youth mental health from 2012 to 2022. An audit of what the evidence actually supports.
Something Changed Around 2012
and it changed in eleven countries at once
Of the eleven OECD countries with continuous time series on youth mental health, nine recorded an annual average decline of between 3% and 16% over the decade from 2012 to 2022 ✓ Established Fact [1]. Only Japan and Korea improved. The synchrony across countries with different economies, school systems and welfare states is the single hardest fact any explanation has to account for, and it is the reason a debate that began as a dispute about screen time has become a dispute about what population-level causal evidence looks like.
The core question is narrow and answerable. Did adolescent mental health deteriorate in a way that began around 2012, and if so, what changed. The first half has an answer that is no longer seriously disputed: the deterioration is measured, it is cross-national, and it predates the pandemic. The OECD, reviewing eleven member countries with comparable time series, found annual average declines in youth mental health status of 3% to 16% between 2012 and 2022, with improvement in only two countries ✓ Established Fact [1]. The second half of the question is where the discipline splits, and it splits along methodological lines that are worth setting out precisely rather than tribally.
Begin with the OECD measurement, because it is the broadest. The 2026 report on child, adolescent and youth mental health describes a marked increase in psychological distress, depressive symptoms and poor mental health among 10 to 24 year olds across the majority of member countries, with few examples of improvement ✓ Established Fact [1]. Crucially, the report treats the deterioration as a long-running trend that predates COVID-19 and intensified afterwards, rather than as a pandemic artefact [1]. That chronology matters: a shock that begins in 2020 would have a simple explanation, and this one does not.
The Health Behaviour in School-aged Children survey, run with the WHO Regional Office for Europe across 44 countries and regions, points the same way from a different instrument. In the 2021/22 round of 279,117 adolescents aged 11, 13 and 15, one third reported feeling nervous or irritable more than once a week over the previous six months, 29% reported sleep difficulties and 25% reported feeling low ✓ Established Fact [3]. Between 2014 and 2022, the share of adolescents reporting frequent low mood and multiple health complaints rose in every HBSC-participating OECD country [2]. Every one.
The United States has the longest and most cited series. The Centers for Disease Control and Prevention report that 40% of American high school students reported persistent feelings of sadness or hopelessness in 2023, against 30% in 2013 ✓ Established Fact [4]. The gender split is wide: 53% of girls against 28% of boys [4]. The 2023 figure is a modest improvement on the 2021 peak of 42% overall and 57% among girls, which is the first genuine sign of a turn in a decade of US data [4]. It is also the detail most often left out of the alarming version of this story.
England provides an independent check with clinical rather than symptom-screen measurement. The Mental Health of Children and Young People survey, commissioned by NHS England and run by the Office for National Statistics with the National Centre for Social Research and the universities of Cambridge and Exeter, found that 20.3% of 8 to 16 year olds had a probable mental disorder in 2023 ✓ Established Fact [8]. The series moved from 12.5% in 2017 to 17.1% in 2020, then held roughly flat through 2023 [8]. A step change followed by a plateau is a different shape from a smooth exponential, and the shape constrains the candidate explanations.
Any national explanation has to survive the comparison test. US opioid deaths, British austerity, Nordic school reform and Australian housing costs are not the same phenomenon, yet the adolescent curves in those countries bend in the same decade and in the same direction. Whatever the cause, it has to be something that arrived in rich countries at roughly the same time and reached adolescents faster than it reached adults. That constraint eliminates most of the standard socioeconomic candidates on its own, which is precisely why the technology hypothesis has proved so durable even among researchers who doubt the size of the effect.
The global burden estimates agree on direction while disagreeing on magnitude. Analyses of the Global Burden of Disease 2021 data find the age-standardised disability-adjusted life year rate for anxiety disorders among 10 to 24 year olds up 18.2% between 1990 and 2021, and for depressive disorders up 13.4%, with a marked acceleration from 2014 onward ◈ Strong Evidence [34]. The fastest rise in depression sits in the 10 to 14 age band, and the increase is largest in high sociodemographic index countries [34]. Rich countries, young adolescents, post-2014: the same coordinates the national surveys produce.
None of this settles causation, and the honest version of the case has to concede three things at the outset. Screening instruments have changed, help-seeking norms have changed, and diagnostic thresholds have moved. Each of those can inflate a measured prevalence without any change in underlying distress. The rest of this report treats those alternatives as live rather than dismissing them, because the strongest version of the sceptical argument is not that nothing happened but that the measurement apparatus moved at the same time as the thing it measures.
Who Moved, and by How Much
girls, internalising symptoms, and the youngest cohort
The aggregate curve hides a sharply asymmetric distribution. The deterioration concentrates in girls, in internalising symptoms such as anxiety and depression rather than externalising ones, and in the youngest adolescents. In the United States the suicide rate for girls aged 10 to 14 rose 13% between 2018 and 2024 while the rate for boys of the same age fell 35%, closing a gender gap that had been nearly two to one as recently as 2017 ✓ Established Fact [7]. Any explanation has to fit that shape, not just the aggregate.
The demographic signature is the most informative feature of the data, and it is remarkably consistent across instruments and countries. The HBSC round found girls reporting worse outcomes than boys on every measure of mental health and well-being, with two thirds of 15 year old girls reporting multiple health complaints ✓ Established Fact [3]. Boys aged 13 and 15 were more likely than girls to report excellent health in nearly every participating country, and the largest gap was among 15 year olds in Denmark, where 43% of boys against 20% of girls reported excellent health [3]. That is not a small difference between sexes. It is a different distribution.
The age gradient runs the same way. Almost twice as many 15 year olds as 11 year olds reported feeling lonely [3]. In the US clinical series, anxiety among adolescents aged 12 to 17 rose from 10.0% in 2016 to 16.1% in 2023, a 61% relative increase, with larger increases among girls and among 12 to 17 year olds than among boys and 6 to 11 year olds ✓ Established Fact [4]. The Global Burden of Disease reanalysis places the fastest depression growth in the 10 to 14 band [34]. Early adolescence, and girls within it, is where the curve bends hardest.
Symptom self-report is vulnerable to changes in reporting norms, so the decisive question is whether hard clinical endpoints moved as well. They did. During February and March 2021, emergency department visits for suspected suicide attempts among US girls aged 12 to 17 ran 51% above the equivalent 2019 period, while the increase among boys of the same age was 4% ✓ Established Fact [6]. Mental-health-related emergency department visits among 12 to 17 year olds rose 31% against 2019 [6]. Those are presentations to hospital, recorded by clinicians, not survey answers.
Between 2021 and 2025, US emergency departments recorded 833,335 visits for suspected suicide attempts, and adolescents aged 12 to 17 accounted for 24.8% of them despite being a small share of the population ✓ Established Fact [5]. Their visit proportion of 82.2 per 10,000 emergency visits was the highest of any age group, and among girls in that band it reached 120.0 per 10,000 [5]. Suicide deaths among US girls aged 10 to 14 more than quadrupled between 2007 and 2024, from 0.51 to 2.28 per 100,000 [7]. Mortality data and hospital presentations are not subject to the prevalence-inflation critique that applies to screening questionnaires.
The mortality series is the most resistant to measurement objections and the most disquieting. Analysis of federal data from 2001 to 2024, published in JAMA Pediatrics, found that by 2024 the suicide rate for US children aged 10 to 14 stood at 2.38 per 100,000 for boys and 2.28 for girls, a gap of roughly 4% ✓ Established Fact [7]. In 2017 boys in that band had nearly twice the rate of girls [7]. The convergence came from both directions: boys down 35% between 2018 and 2024, girls up 13% [7]. The narrowing appeared among Black, Hispanic and White children alike [7].
The most recent US surveillance data complicate the straightforward decline narrative, and the complication deserves equal billing. The CDC reported in June 2026 that emergency department visit proportions for suspected suicide attempts fell 7.0% overall between 2021 and 2025, and by 20.8% among 12 to 17 year olds, with girls in that band down 22.3% ✓ Established Fact [5]. State-level Youth Risk Behavior Surveys from the 2025 cycle point the same way, with New Hampshire recording 34% of students reporting persistent sadness against 44.2% in 2021 [4]. Adolescent distress may have peaked.
Against the hard endpoints stands the most serious sceptical argument, which is not that the data are fabricated but that awareness itself changes prevalence. Lucy Foulkes and Jack Andrews set out the prevalence inflation hypothesis in 2023: mental health awareness campaigns may raise reported rates through improved recognition of genuine problems, and separately through overinterpretation, in which ordinary distress is relabelled as illness ⚖ Contested [27]. The second mechanism can become self-fulfilling if relabelling changes how adolescents interpret and respond to normal difficulty [27]. School-based programmes teaching cognitive behavioural principles have in some trials increased internalising symptoms relative to controls [27].
The prevalence inflation hypothesis explains part of the self-report series well and the mortality series badly. Suicide deaths are counted by coroners, not by adolescents describing their feelings. If awareness effects were doing all the work, the questionnaire curves and the mortality curves would have diverged, and in the United States they did not. The most defensible reading is that both processes are running at once: a real deterioration in the distribution of adolescent distress, amplified in the survey instruments by a genuine change in how young people label what they feel.
The Window Between 2010 and 2015
what actually arrived, and when
The case for 2012 as an inflection rests on a coincidence of adoption curves. In early 2012, 23% of US teenagers aged 12 to 17 said their phone was a smartphone; by 2013 it was 37%, up from 23% in 2011 ✓ Established Fact [28]. Facebook bought Instagram in April 2012, front-facing cameras had shipped two years earlier, and the algorithmic feed replaced the chronological one across the same window. The question is whether that clock ran closely enough to the mental health clock to carry evidential weight, and whether other clocks ran alongside it.
Adoption data set the boundaries of the argument. Pew Research Center found that 46% of American adults owned a smartphone as of February 2012, up eleven percentage points in nine months, and that ownership had reached a tipping point at which smartphones rather than feature phones dominated new purchases ✓ Established Fact [28]. Among teenagers the transition was slightly later and considerably faster: 23% in 2011, 37% by early 2013, with 31% of 14 to 17 year olds already on smartphones in 2012 against 8% of 12 to 13 year olds [28]. The device reached older adolescents first.
Three product changes compounded the hardware shift inside the same five-year window. Front-facing cameras shipped on mainstream handsets from 2010, which made the self-portrait the default unit of social exchange. Instagram launched in October 2010 and was acquired by Facebook in April 2012, after which its growth accelerated sharply [28]. And the major platforms moved from chronological to ranked feeds, which converted social media from a messaging medium into a comparison medium optimised by engagement metrics. It is the comparison mechanism, not screen time as such, that the causal literature keeps returning to. By 2026, British 13 and 14 year olds were spending an average of 4 hours 19 minutes a day online, and 33% of child social media users reported feeling pressure to be popular online all or most of the time [26].
The timing argument has real force, and it also has a well-known weakness. Several other things changed in rich countries across the same decade. The aftermath of the 2008 financial crisis reshaped household security and youth employment prospects. Academic competition intensified, sleep duration fell, unsupervised outdoor time contracted, and in the United States the opioid epidemic reorganised family structures in specific regions. The OECD explicitly lists digitalisation and social media alongside climate anxiety, fears about global conflict, socioeconomic pressure, bullying, academic stress, inequality and poverty as intersecting drivers ⚖ Contested [2].
Jonathan Haidt has pressed the timing objection in the opposite direction, and his version deserves a fair statement. Adolescent mental illness rates were flat through the 2000s and turned upward roughly four years after the 2008 crisis, during the recovery rather than the downturn [30]. Depression rose fastest among higher-income teenagers after 2012, which is the opposite of what an economic-hardship model predicts [30]. And the deterioration appears in Nordic countries that lack the specific American pathologies of opioids, firearm exposure and mass incarceration [30]. Each point is a genuine constraint on the alternatives.
Smartphone adoption, algorithmic feeds, declining teen sleep, falling unsupervised play and post-crisis economic anxiety all accelerate in the same five-year window across the same group of countries. That is the analytical trap at the centre of this dispute. When several candidate causes share a time series, aggregate timing evidence cannot separate them, and the debate can only be resolved by designs that create variation in one candidate independently of the others. That is exactly what the broadband rollout and platform rollout studies attempt, and why they carry more weight than any number of cross-sectional correlations.
The Japanese case is the most instructive anomaly and it cuts both ways. Japan is one of the two OECD countries where the youth mental health indicator improved across the 2012 to 2022 decade [1]. Yet over the same period, school non-attendance rose without interruption. In the 2024 school year, 353,970 Japanese elementary and junior high school students were recorded as futoko, absent for 30 days or more for psychological, emotional or social reasons, the twelfth consecutive annual increase and more than double the figure a decade earlier ✓ Established Fact [29].
The Japanese figures amount to 3.86% of all students, 2.30% in elementary and 6.79% in junior high, with roughly 67,000 absent for 90 days or more [29]. The ministry attributes 51.8% of cases to apathy or anxiety [29]. Whether that counts as deteriorating adolescent mental health depends entirely on which indicator is treated as the outcome. The disagreement between Japan improving on one measure and setting records on another is a warning about how much of this debate is driven by instrument choice rather than by the underlying phenomenon.
The Case That Phones Did It
dose, timing, synchrony and internal documents
The strongest version of the causal argument does not rest on correlation. It rests on a stack: dose-response gradients in large surveys, within-person longitudinal designs in which the arrow runs from use to symptoms, randomised restriction experiments with small but consistent effects, quasi-experimental rollout studies using hospital records, and internal company research now entering the public record through litigation ◈ Strong Evidence [19]. Each layer is individually contestable. The claim is that the pattern across layers is not.
Haidt and Zachary Rausch set out the case at institutional scale in the World Happiness Report 2026, in a chapter titled with the claim itself: social media is harming adolescents at a scale large enough to cause changes at the population level [19]. The argument proceeds from testimony, through the academic literature on causal impact, to the policy conclusion. The chapter sits in the same volume as a separate analysis by Jean Twenge and colleagues using OECD survey data, and the pairing is deliberate: the authors are arguing that converging methods now point the same way [20].
The dose-response evidence is the oldest layer and the weakest on its own. A meta-analysis of 26 studies found depression risk rising by roughly 13% for each additional hour of daily social media use, and a 2018 study of 14 year olds found girls using social media five or more hours a day around three times as likely to be depressed as light users [30]. Dose gradients are consistent with causation and equally consistent with distressed adolescents using more social media. On their own they establish nothing, which Haidt concedes.
Twenge and colleagues sharpened the dose picture using the OECD Programme for International Student Assessment, which in 2022 asked more than 270,000 fifteen and sixteen year olds across 47 countries about daily social media hours and life satisfaction ◈ Strong Evidence [20]. Among girls, mean life satisfaction peaked among light users of under an hour a day and declined with every additional hour [20]. Among boys the same pattern held only in Western Europe and the English-speaking countries [20]. The heaviest users of seven hours or more also showed the widest dispersion, clustering at both the top and the bottom of the life-satisfaction scale [20].
Social media is harming adolescents at a scale large enough to cause changes at the population level.
— Jonathan Haidt and Zachary Rausch, World Happiness Report 2026, March 2026The experimental layer is where the argument becomes testable. Haidt counts 22 experimental studies of social media reduction, of which 16 found significant evidence of harm from use [30], and 8 of 9 quasi-experiments finding mental health damage concentrated among girls [30]. The best-known experiment deactivated Facebook accounts for 2,743 adults and found improved well-being, with 80% of the treatment group agreeing afterwards that deactivation had been good for them [30]. A separate trial combining social media reduction with physical exercise produced the largest decreases in depressive symptoms [30].
The within-person longitudinal evidence is the newest layer and the most methodologically pointed, because it addresses the reverse-causation objection directly. Jason Nagata and colleagues analysed 11,876 participants in the Adolescent Brain Cognitive Development study across four annual waves using random-intercept cross-lagged panel models ◈ Strong Evidence [12]. Increases in an individual adolescent's social media use above that same adolescent's personal average predicted greater depressive symptoms a year later, with a standardised coefficient of 0.07 from year one to year two and 0.09 from year two to year three [12].
Random-intercept cross-lagged panel models separate within-person change from stable between-person differences, which is what makes the Nagata analysis harder to dismiss than ordinary longitudinal correlation ◈ Strong Evidence [12]. Across 11,876 children entering the cohort at a mean age of 9.9 years, within-person increases in social media use predicted greater depressive symptoms one year later at both transitions tested, while depressive symptoms did not predict subsequent increases in social media use [12]. The effect sizes are small. The asymmetry in direction is the finding.
The restriction trials produce a small but non-zero average effect, and the exact magnitude is itself a live methodological fight. A 2025 meta-analysis of randomised controlled trials of social media restriction estimated an average effect on overall well-being of 0.17 standard deviations, with a confidence interval from 0.09 to 0.25, after removing effect sizes the authors judged problematic ◈ Strong Evidence [17]. That is small by conventional benchmarks. Whether an effect of that size at the individual level can produce a visible population-level shift depends on how many adolescents are exposed and for how long, which is an arithmetic question rather than a psychological one.
The final layer is documentary rather than statistical, and it is arriving through discovery. Internal Meta research surfaced through whistleblower disclosure and state litigation indicates that teenagers themselves attributed rises in anxiety and depression to Instagram, unprompted and consistently [30]. Tennessee filings describe a May 2020 internal presentation of 97 pages synthesising adolescent development, neuroscience and nearly 80 pieces of the company's own product research [31]. Whether internal awareness of harm constitutes evidence of harm is a legal question more than a scientific one, but it bears on intent and therefore on regulation.
The Case That It Did Not
effect sizes, confounders and the reverse arrow
The sceptical position is not that adolescents are fine. It is that the effect attributed to social media is too small, too inconsistent across datasets and too confounded to carry the causal weight placed on it. The National Academies of Sciences, Engineering, and Medicine reviewed the literature in 2024 and concluded that it did not support the conclusion that social media causes changes in adolescent health at the population level ⚖ Contested [14]. A Swedish national cohort followed to 2026 found associations that disappeared entirely once confounders were controlled [18].
Candice Odgers set out the sceptical case most directly in a review of The Anxious Generation for Nature in March 2024. Her argument is not that adolescent distress is imagined but that the specific mechanism the book proposes is unsupported: the suggestion that digital technologies are rewiring children's brains and causing an epidemic of mental illness is not, on her reading, backed by the evidence ⚖ Contested [13]. She points to an analysis across 72 countries finding no consistent or measurable association between well-being and the global rollout of social media [13].
Odgers also cites the Adolescent Brain Cognitive Development study, the largest long-term study of adolescent brain development in the United States, which has found no evidence of drastic neurological changes associated with digital technology use [13]. Her alternative explanation runs through structural causes: racism, poverty, the opioid epidemic and the long aftermath of the 2008 financial crisis [13]. And she proposes a different arrow entirely, in which adolescents who are already struggling reach for their phones in much the way struggling adults do [13].
The institutional weight behind that position is substantial. The National Academies committee report of 2024 examined the difficulty of establishing causality in complex social phenomena and declined to endorse population-level causation [14]. It did identify plausible harm pathways, including unhealthy social comparison and the displacement of sleep, exercise and study time, noting that sleep loss is itself a risk factor for depression, mood disturbance, injury, attention problems and weight gain [14]. Its recommendation was cautious intervention and further research into specific platform features rather than broad bans [14].
The book's repeated suggestion that digital technologies are rewiring our children's brains and causing an epidemic of mental illness is not supported by science.
— Candice Odgers, review of The Anxious Generation, Nature, March 2024The largest single dataset brought to the question points away from harm. Matti Vuorre and Andrew Przybylski analysed Gallup World Poll responses from 2,414,294 individuals aged 15 to 99 across 168 countries between 2006 and 2021, covering eight measures of well-being including life satisfaction, social life, purpose and community ◈ Strong Evidence [15]. People with internet access or active internet use reported meaningfully greater well-being than those without [15]. The authors concluded that two decades of data showed only small and inconsistent changes in global well-being, not the pattern a technology-driven collapse would produce [15].
The newest long-run cohort evidence is the most awkward for the causal case. Martina Zetterqvist and colleagues followed a nationally representative Swedish cohort, surveying 5,535 ninth-grade students aged 15 or 16 in 2017 and reassessing 3,193 of them in 2022, measuring depression and anxiety with the Patient Health Questionnaire-4 ⚖ Contested [18]. In unadjusted models, adolescent social media hours predicted anxiety and depression five years later [18]. After adjustment for confounders, the associations were no longer statistically significant, with no effect modification by gender, socioeconomic status or physical activity [18].
The Swedish national cohort published in the Journal of Adolescent Health in July 2026 is the strongest test yet of the long-run version of the claim, because it follows the same adolescents from age 15 or 16 into their early twenties ⚖ Contested [18]. The unadjusted association between social media hours and later anxiety and depression was positive and significant, which is what almost every cross-sectional study reports [18]. It did not survive controlling for the characteristics that predict both heavy use and poor mental health [18]. That is the pattern of a confounded association rather than a causal one.
The experimental meta-analytic picture is contested to the point of being a dispute about the meta-analyses themselves. Christopher Ferguson published a preregistered meta-analysis of social media experiments finding no average effect on mental health [16]. It was then criticised in detail over study inclusion and exclusion decisions, averaging across heterogeneous experiment types, reliance on a summary outcome combining unlike mental health domains, alleged inconsistencies in sample and effect size calculation, and absence of moderation analysis by intervention length [17]. The subsequent restriction meta-analysis found a small significant effect [17].
What the sceptics are pressing is a question about magnitude rather than existence. A within-person coefficient of 0.07 and a restriction effect of 0.17 standard deviations are real but modest, and the inferential leap from those numbers to a population-level epidemic requires assumptions about exposure breadth and duration that are rarely stated explicitly. The counter-argument is arithmetic: a small effect applied to almost every adolescent in the developed world, for several hours a day, over a decade, is not obviously incapable of moving a population distribution.
The Strongest Designs
broadband rollouts, campus Facebook and hospital records
Quasi-experiments exploit variation in technology arrival that is unrelated to mental health trends, which is what allows them to support causal claims that correlation cannot. Three of them use administrative or clinical records rather than questionnaires. The Italian broadband study finds mental disorders rising 0.08 standard deviations among cohorts exposed before the age of 20 and no effect at all on older cohorts ◈ Strong Evidence [10]. The Spanish fibre study finds effects on girls and none on boys [11]. The campus Facebook study finds increased use of mental healthcare among susceptible students [9].
The design logic is worth stating plainly, because it is what separates this literature from the rest. If broadband arrives in a town because that town sits close to a telephone exchange built decades before the internet existed, then the timing of arrival is unrelated to that town's adolescent mental health trajectory. Comparing outcomes across towns with different arrival dates then approximates an experiment. The same logic applies to a platform that expanded college by college on a schedule driven by institutional characteristics rather than student psychology.
Luca Braghieri, Roee Levy and Alexey Makarin applied that logic to the staggered introduction of Facebook across US colleges, published in the American Economic Review in November 2022 ◈ Strong Evidence [9]. Using a generalised difference-in-differences design against student mental health data, they found that the arrival of Facebook at a college worsened student mental health and increased reported academic impairment attributed to poor mental health [9]. Among students predicted to be most susceptible to mental illness, Facebook arrival increased utilisation of mental healthcare services [9]. The mechanism evidence pointed to unfavourable social comparison [9].
Dante Donati, Ruben Durante, Francesco Sobbrio and Dijana Zejcirovic ran the strongest of the administrative-records designs. Using Italian hospital data on diagnosed mental disorders from 2001 to 2013 combined with municipal broadband availability, they found that broadband access raised the prevalence of mental disorders among cohorts born between 1985 and 1995 by 0.08 standard deviations, with no effect on those born between 1974 and 1984 ◈ Strong Evidence [10]. The effect was driven by individuals exposed before the age of 20 [10].
The Italian study is difficult to explain away as increased detection, because the effects persist when the analysis is restricted to self-harm and to urgent or compulsory hospitalisations, which are not discretionary presentations ◈ Strong Evidence [10]. The affected diagnoses span depression, anxiety, drug abuse and personality disorders in both sexes, with eating disorders additionally affected among females [10]. The age-of-exposure boundary is the striking result: the cohort born between 1974 and 1984, which received the same broadband at an older age, shows no effect at all [10].
The Spanish evidence supplies the gender specificity the epidemiological data demand. Esther Arenas-Arroyo, Daniel Fernandez-Kranz and Natalia Nollenberger exploited the rollout of fibre-to-the-home across Spanish provinces between 2007 and 2019, measured against hospital discharge diagnoses for behavioural and mental health conditions among adolescents ◈ Strong Evidence [11]. The effect was positive and significant for girls and absent for boys [11]. The mechanism evidence showed high-speed internet increasing addictive internet use and reducing time spent sleeping, doing homework and socialising with family and friends, with girls driving all of those shifts [11].
Three studies, three countries, three data sources, one signature. The effects concentrate in adolescents rather than adults, in girls rather than boys, in clinical presentations rather than self-report, and in displacement of sleep and social contact rather than in anything exotic. Haidt counts 8 of 9 quasi-experiments finding mental health harm concentrated among girls [30]. That convergence is the part of the causal case that the sceptical literature has engaged with least, and it is the part that deserves the most weight.
Broadband is not social media. Fibre-to-the-home delivers gaming, pornography, streaming, gambling and social platforms simultaneously, and none of these designs can attribute the measured effect to one of them. The Italian data end in 2013, before the algorithmic feed became standard, and the campus Facebook study observed a product that no longer exists in the form studied. These are estimates of what arriving high-bandwidth internet did to adolescents in a particular decade, which is a narrower and more defensible claim than the one usually made on their behalf.
The limits run further than product identification. Every one of these designs measures the effect of exposure on those who received it earlier rather than later, which says nothing about what would happen if the technology were withdrawn from a population that has already restructured its social life around it. The restriction trials attempt that question directly and return the small effects described earlier [17]. Removal and non-arrival are different interventions with different expected magnitudes, and conflating them is the most common error in the policy discussion.
The honest summary of the quasi-experimental literature is that it establishes a real, modest, age-specific and sex-specific harm from high-bandwidth internet arrival during adolescence, measured in clinical outcomes. It does not establish that social media specifically caused the post-2012 inflection, and it does not quantify how much of the observed population change any such mechanism would explain. That is a considerably stronger conclusion than the sceptics allow and a considerably weaker one than the policy response assumes.
Policy Has Outrun the Evidence
bans arrived before the trials finished
Australia removed 4.7 million under-16 accounts within weeks of its ban taking effect on 10 December 2025 ✓ Established Fact [24]. By August 2026, roughly 70% of Australian under-16s were still reported by their parents to be reaching social media [25]. Denmark, the United Kingdom and the European Union are moving along the same path. Meanwhile the best-designed evaluation of school phone bans found no mental wellbeing benefit at all [22], and the best-designed evaluation of a national ban found large ones [21]. The policy is ahead of the science in both directions.
Australia legislated first and fastest. The Online Safety Amendment (Social Media Minimum Age) Act 2024 took effect on 10 December 2025, requiring major platforms to take reasonable steps to prevent under-16s from holding accounts, covering Instagram, Facebook, TikTok, Snapchat, X and YouTube ✓ Established Fact [24]. By mid-December, age-restricted platforms had removed access to 4.7 million accounts judged to belong to under-16s, and the government confirmed the figure in mid-January 2026 [24]. It was the first national age limit of its kind anywhere.
Enforcement has proved considerably harder than removal. On 31 March 2026 the eSafety Commissioner opened formal investigations into Facebook, Instagram, Snapchat, TikTok and YouTube over suspected breaches, the first public compliance assessment since the Act commenced [24]. The central evidence was a survey finding that approximately 70% of Australian under-16s were still reported by their parents to be accessing social media after the ban [25]. Removing an account and preventing access turn out to be different problems, and only the first is within a platform's straightforward technical control.
Europe is following on a staggered schedule. Denmark agreed to restrict social media for under-15s, with parental override available for 13 and 14 year olds, enforcement leaning on the national electronic identity system and a dedicated age-certificate application, and the measure potentially becoming law by mid-2026 [33]. The United Kingdom brought Online Safety Act age assurance duties into force on 25 July 2025, and providers processed 5.7 million age checks on the first day [32]. By February 2026 Ofcom had opened investigations into more than 90 services and issued six fines [32].
School phone policy is the one area with genuine evaluation evidence, and it is contradictory. Sara Abrahamsson studied the staggered introduction of smartphone bans across Norwegian middle schools and found that girls made almost 60% fewer visits to psychological specialists for mental health issues and around 29% fewer consultations with general practitioners ◈ Strong Evidence [21]. Girls' average grades and mathematics test results improved, bullying fell, and the largest gains accrued to girls from lower socioeconomic backgrounds [21]. There was no effect on the likelihood of being diagnosed or treated [21].
The best-powered English evaluation found the opposite. The SMART Schools study compared 1,227 adolescents aged 12 to 15 across 30 schools with restrictive and permissive phone policies and found no evidence that restriction was associated with better mental wellbeing, better sleep or better educational outcomes ◈ Strong Evidence [22]. Restrictive policies reduced in-school phone use by roughly 40 minutes but did not reduce overall daily phone or social media use [22]. The authors concluded that the evidence does not support school phone prohibition in its current form [22]. Displacement, not reduction.
| Risk | Severity | Assessment |
|---|---|---|
| Policy locked in ahead of the evidence base | Australia, Denmark and the United Kingdom are committing to age-based exclusion while the best evaluation of school restriction found no wellbeing effect [22] and the longest cohort study found no surviving association [18]. Reversing a national identity-verification regime is far harder than introducing one. | |
| Age assurance creates a new identity-data surface | Enforcement requires verifying the age of every user, not only minors. Britain processed 5.7 million checks on the first day of its regime [32] and Denmark plans to route verification through the national electronic identity system [33]. The privacy exposure is systemic and permanent. | |
| Displacement rather than reduction | The SMART Schools evaluation found in-school restriction cut phone use during the school day without reducing total daily use [22], and roughly 70% of Australian under-16s were still reaching social media after the ban [25]. Restriction that moves behaviour rather than reducing it delivers none of the hypothesised benefit. | |
| Attention diverted from demonstrated drivers | The OECD lists bullying, academic stress, inequality, poverty, climate anxiety and conflict fears alongside digitalisation [2], and the National Academies identify sleep displacement as a specific pathway [14]. A single-cause policy frame risks defunding interventions with better evidence. | |
| Erosion of the measurement infrastructure | The entire debate rests on a small number of long-running surveys. Disruption to the Youth Risk Behavior Survey and comparable instruments would remove the only means of evaluating whether any of these interventions worked [35], leaving policy permanently unfalsifiable. |
The Dutch evaluation sits between the two and is the weakest design of the three. After the national classroom device ban took effect on 1 January 2024, a government-commissioned survey of 317 secondary schools found three quarters reporting positive effects on student concentration, nearly two thirds reporting improved social climate and about one third reporting better academic performance [23]. These are school-reported perceptions rather than measured outcomes, which places them below the Norwegian administrative-records design and the English controlled comparison in evidential weight.
The courts may resolve what the regulators cannot. A coalition of 29 US states opened trial against Meta in Oakland on 18 August 2026, the first federal social media addiction case to reach a jury, with California, Colorado, Kentucky and New Jersey going first [31]. Six school district bellwether cases alleging public nuisance are scheduled for February 2027 [31]. Litigation discovery has already produced more internal platform research than a decade of voluntary disclosure, and that is likely to remain the most productive evidence channel.
What the Evidence Actually Supports
a narrower claim that survives scrutiny
Strip the argument down to what survives adversarial review and a defensible position remains. Adolescent mental health deteriorated across rich countries from roughly 2012, concentrated in girls and in early adolescence, and it appears in mortality and hospital data as well as in questionnaires ✓ Established Fact [1]. High-bandwidth internet arrival during adolescence causes measurable clinical harm, concentrated in girls [11]. What does not survive is the claim that this mechanism explains most of the population change, or that removing it would reverse the curve.
The most useful move at this point is to separate the four claims that are routinely bundled together. First, that adolescent mental health deteriorated. Second, that the deterioration began around 2012 and is cross-national. Third, that smartphones and social media caused it. Fourth, that restricting adolescent access will reverse it. The first two are established. The third is partially supported for a narrower mechanism than usually stated. The fourth has the weakest evidence of all and is the one currently being legislated.
On the first two claims the dispute is effectively over. The OECD documents declines in nine of eleven countries with time series [1], HBSC records rises in low mood and health complaints in every participating OECD country between 2014 and 2022 [2], the English series steps from 12.5% to 17.1% between 2017 and 2020 [8], and US suicide mortality among girls aged 10 to 14 more than quadrupled from 2007 to 2024 ✓ Established Fact [7]. Sceptics of the causal claim do not generally dispute this description, and Odgers herself does not [13].
On the third claim the quasi-experimental evidence supports a specific and limited version. Broadband arrival raised diagnosed mental disorders among cohorts exposed before age 20 by 0.08 standard deviations in Italy [10], fibre rollout raised adolescent hospital discharges for mental health conditions among Spanish girls and not boys [11], campus Facebook arrival worsened student mental health and raised mental healthcare use among susceptible students [9], and within-person increases in use predicted later depressive symptoms in the largest US adolescent cohort while the reverse did not hold ◈ Strong Evidence [12].
What the evidence supports
Nine of eleven OECD countries with time series declined annually between 2012 and 2022, and every HBSC-participating OECD country recorded rising low mood between 2014 and 2022 [1].
US girls aged 12 to 17 present at emergency departments for suspected suicide attempts at 120.0 per 10,000 visits, and the suicide gender gap among 10 to 14 year olds has nearly closed [7].
Mortality records, emergency presentations and compulsory hospitalisations all move, which the prevalence-inflation explanation cannot account for [5].
Three independent quasi-experiments using administrative records find effects concentrated in adolescents, in girls, and in cohorts exposed before the age of 20 [10].
Random-intercept cross-lagged models across 11,876 adolescents found use predicting later depressive symptoms, with no reverse effect detectable [12].
What the evidence does not support
The rollout designs measure high-bandwidth internet as a bundle and cannot separate social platforms from gaming, streaming, pornography or gambling [10].
Restriction trials average 0.17 standard deviations on well-being and within-person coefficients sit at 0.07 to 0.09, magnitudes conventionally described as small [17].
The Swedish national cohort followed to 2026 found the five-year association between adolescent social media hours and later anxiety and depression vanished after controlling for confounders [18].
The SMART Schools comparison found restrictive school policies reduced in-school use by 40 minutes with no wellbeing, sleep or attainment benefit [22].
The OECD identifies digitalisation alongside climate anxiety, conflict fears, socioeconomic pressure, bullying, academic stress and inequality as intersecting drivers [2].
The fourth claim is where the asymmetry between evidence and action is starkest. The two rigorous evaluations of school restriction disagree: the Norwegian administrative-records study found large reductions in girls' specialist mental health visits [21], and the English controlled comparison found none [22]. The Norwegian design has better outcome data, the English design has better exposure comparison, and neither has been replicated. Legislating national exclusion on that evidence base is a policy judgement about precaution, not an inference from data, and it should be described as such.
There is a defensible precautionary case, and it does not require winning the causal argument. The intervention is reversible where a generation is not. The harms of restriction fall mainly on convenience and on platform revenue, while the harms of inaction, if the causal case is right, fall on adolescent mortality. What that reasoning cannot license is the claim that the science is settled, because the strongest sceptical evidence, including the 2026 Swedish cohort and the Odgers reading of the 72-country analysis, is recent and methodologically serious.
Australia, Denmark and the United Kingdom have now created the natural experiment this field has lacked. From December 2025, one high-income country with excellent health administrative data removed social media access from a defined age cohort while comparable countries did not. If adolescent emergency presentations, self-harm admissions and suicide mortality in Australian 13 to 15 year olds diverge from the New Zealand and Canadian series over the next three years, the causal question is close to settled. If they do not, that is equally informative. Preserving the measurement instruments matters more than the ban itself.
Three research priorities follow from the current state of evidence. First, evaluate the Australian cohort against matched comparison countries using administrative health records rather than surveys, which requires funding the linkage now rather than after the fact. Second, run restriction trials long enough to detect cumulative effects, since every existing trial is short relative to a decade of exposure. Third, separate the bundle, because the rollout designs measure high-bandwidth internet and the policy targets social platforms, and nobody has yet shown those are the same thing.
The measured position is less satisfying than either headline. Adolescent mental health deteriorated across the rich world from the early 2010s, girls bore most of it, and the deterioration is visible in mortality as well as in questionnaires. High-bandwidth internet arriving during adolescence causes real clinical harm, most clearly in girls, at a magnitude that is modest per person and applied to nearly everybody. Whether that accounts for most of the inflection remains genuinely open, and the countries now legislating have made the next decade of data the decisive evidence.