The teen phone debate

A considered reading of fifteen years of research on phones, social media, and adolescent mental health — what is actually known, what is seriously contested, and what the loudest commentators on each side have overclaimed.

---

Of all the strands of research touched on by this site, none has attracted more public attention, more confident assertion, and more methodological scrutiny than the question of what smartphones and social media have done to teenagers. None is more contested. None is more often misread.

Jonathan Haidt's 2024 book The Anxious Generation argues, in essence, that the rapid adoption of the smartphone and the visual social platforms that ride on it — Instagram, Snapchat, TikTok — caused a measurable, generation-defining decline in adolescent mental health beginning around 2012. The decline is largest in girls. The mechanism, in his account, is the rewiring of childhood: a play-based culture supplanted by a screen-based one in a five-year window between 2010 and 2015. The book has been a bestseller, a policy catalyst, and, depending on which researcher you ask, either a long-overdue alarm or a confident overreach (Haidt, 2024).

A near-equally prominent group of researchers — Candice Odgers at UC Irvine, Amy Orben at Cambridge, Andrew Przybylski at Oxford, Matti Vuorre at Tilburg, Sonia Livingstone at the LSE — have argued, also at length, also in print, that the empirical record does not support the scale of Haidt's claim. They do not deny that something has changed in the lives of teenagers. They do not deny that some adolescents are harmed by some forms of online use. They argue, with the data behind them, that the average effects across the adolescent population are small, the strongest causal evidence is weak, the methodology has been frequently overstretched, and the most useful interventions are not the ones the alarm suggests (Odgers, 2024; Mansfield et al., 2025).

Both sides have read the same papers. Both sides include serious researchers. The disagreement is not about basic competence; it is about how to weight imperfect evidence, where to draw the line between a real effect and a noise floor, and whether mean effects across a population are the right object of inquiry when the population is heterogeneous. The way the public discourse handles this disagreement is mostly to take one side and dismiss the other. The literature itself is more interesting than that.

What follows is an attempt at the more interesting reading. The strongest claims on each side are presented as steelman positions before the critiques are introduced. Confidence tiers, where they matter, are flagged explicitly. Where the evidence is contested, that is how it is described.

The chronological case

The first thing to understand about the Twenge–Haidt position is that its strongest evidentiary leg is a timing argument, not a mechanism argument. Beginning around 2012 and accelerating through the mid-2010s, multiple independent measures of adolescent mental health in the United States, the United Kingdom, Canada, Australia, and several Nordic countries showed sharp inflections. Depression diagnoses rose. Self-harm presentations to emergency departments rose. Suicide rates among girls aged 10 to 14 rose. Self-reported loneliness, anxiety, and unhappiness rose. The trends were largest among girls, and were not present at anything like the same magnitude in adolescents over twenty or in adults more broadly.

Jean Twenge, a psychologist at San Diego State University, has documented these inflections across multiple papers and datasets. Her 2020 review Increases in Depression, Self-Harm, and Suicide Among U.S. Adolescents After 2012 is a useful summary of the trend data and the proposed mechanisms (Twenge, 2020). The phenomenon she describes is not a small drift; it is a step change visible in nationally representative samples and in clinical data, with consistent timing across the populations measured. That such a change occurred is not, in the current literature, seriously disputed.

A second line of evidence concerns whether the same change is visible internationally — and if so, whether it tracks something other than US-specific factors like recession recovery, opioid spillover, or the post-2012 escalation in school shootings. Twenge and Haidt, with collaborators, analysed the OECD's Programme for International Student Assessment loneliness measure, included in the PISA waves of 2000, 2003, 2012, 2015, and 2018. The dataset includes over a million students aged 15 to 16 across 37 countries. Between 2012 and 2018, school loneliness rose in 36 of those 37 countries. The pattern was larger in girls than in boys. Multi-level models found that loneliness was elevated in years and countries with high smartphone access and high internet use; it was lower in countries with high unemployment rates. Other macroeconomic variables — GDP, income inequality, family size — did not track the rise (Twenge et al., 2021).

A third strand involves what happens when you separate light from heavy users on the same self-report measures. Twenge and Campbell (2019), analysing three large adolescent surveys totalling 221,096 respondents, found that adolescents who used digital media five or more hours per day were between 48% and 171% more likely to report depression, suicidal ideation, or past suicide attempts than adolescents who used digital media less than one hour per day. The heavy-user group was roughly twice as likely to report a past suicide attempt. One caveat the paper itself acknowledges: the surveys aggregate diverse activities — social media, gaming, television, texting — into a single digital media bucket, which makes it harder to isolate the contribution of social media specifically. The discontinuity between moderate use (one to two hours) and heavy use (five-plus) is the load-bearing finding; the dose–response is non-linear, with the bulk of the effect concentrated in the upper tail.

Haidt's book pulls these strands together with a fourth: an account of mechanism. The proposal is what he calls the great rewiring — the period between 2010 and 2015 in which the smartphone became universal among American teenagers, the front-facing camera became standard, the platforms became algorithmically curated, and the play-based childhood of previous generations was supplanted by a screen-based one in which most peer interaction happened through a device. The four core harms Haidt names are sleep deprivation, attention fragmentation, social deprivation, and addictive reinforcement. He is willing to argue, on the basis of the timing and the cross-national consistency, that the rewiring is causally connected to the mental health inflection (Haidt, 2024).

This is the strongest version of the case. It is not the only version: Twenge has been more cautious in her academic writing than Haidt has been in his book, and the two voices should not be flattened together. But the Haidt version is the one most readers encounter, and the version the popular discussion is structured around. Its strengths are the consistency of timing across populations, the size of the heavy-user effects in cross-sectional data, and the plausibility of multiple converging mechanisms. Its weaknesses, which the next section treats at length, concern the gap between the population-level inflection and individual-level causation.

The small-effect case

The counter-position is not that the population-level inflection isn't real. The leading critics — Odgers, Orben, Przybylski, Vuorre, Livingstone — accept the depression and self-harm trend data. The disagreement is about whether those trends can be reliably attributed to phones and social media when the most rigorous individual-level studies find such small associations.

The flagship paper on this side of the debate is Orben and Przybylski's 2019 Nature Human Behaviour paper The association between adolescent well-being and digital technology use. Its methodological contribution was an analytic technique called specification curve analysis: instead of choosing one analytic configuration and reporting it, the authors ran the same question across hundreds of theoretically defensible combinations of variable coding, control variables, and outcome measures, then reported the distribution of effect sizes. Across three large datasets totalling 355,358 adolescents, the median association between digital technology use and well-being was negative but very small — explaining at most 0.4% of the variation in well-being (Orben & Przybylski, 2019a).

The comparison that became famous was their illustration of how that effect compared to other variables in the same datasets. The association was small enough that some neutral comparison variables looked similar or larger: eating potatoes was nearly as negative as technology use in one dataset, and wearing glasses was more negative in another. By contrast, being bullied was several times more strongly associated with lower well-being than technology use was. The point was not that potatoes matter; it was that a 0.4% variance figure should not bear the policy weight that headlines had been placing on it.

Orben and Przybylski followed this up with a paper in Psychological Science that addressed the dominant methodological weakness in the field — that screen-time questions on surveys are inaccurate — by using time-use diaries instead. Across three nationally representative datasets from Ireland, the United States, and the United Kingdom (N = 17,247), they found little evidence for substantial negative associations between digital-screen engagement and adolescent well-being. The signal that did appear in self-report measures attenuated when digital-screen engagement was measured by time-use diary (Orben & Przybylski, 2019b). A subsequent paper specifically on sleep, using time-use diaries on 11,884 UK adolescents from the Millennium Cohort, found that an additional hour of digital screen time per day predicted only a nine-minute reduction in total sleep on weekdays, and three minutes on weekends; screen use in the half-hour before bed predicted a one-minute reduction (Orben & Przybylski, 2020).

The methodological argument is straightforward. Surveys ask adolescents to estimate their phone use; the estimates are wrong, often by hours; the resulting associations with depression scores describe the overlap between beliefs about phone use and depression, not phone use and depression. When the estimates are replaced by logged or diary-based measurement, the effects shrink. In the time-use studies, they shrink to the point of practical inconsequence.

A second strand of the Oxford–Cambridge counter-position uses cross-national variation as a test. Vuorre and Przybylski (2023) examined Facebook adoption rates in 72 countries between 2008 and 2019 against well-being measures for 946,798 individuals from the Gallup World Poll. If the platform were doing population-level harm, the prediction is that countries where Facebook penetration rose faster should show steeper declines in well-being. They found, if anything, the opposite: Facebook adoption predicted life satisfaction and positive experiences positively, not negatively, both between countries and within countries over time. The associations were small and confidence-bounded; they did not run in the direction the harm hypothesis predicts. A 2024 paper extending the analysis to general internet adoption across a broader sample of countries from 2006 to 2020 reached a similar conclusion (Vuorre & Przybylski, 2024).

In 2025, Mansfield, Ghai, Hakman, Ballou, Vuorre, and Przybylski published a methodological critique in The Lancet Child & Adolescent Health that summarised the position. The argument: the field has been mis-served by underpowered cross-sectional designs with self-report measures, by inappropriate aggregation across heterogeneous platforms and uses, by overconfident extrapolation from null or small findings, and — uniformly — by activists on both sides reading more into the data than the methodology supports (Mansfield et al., 2025). The piece is worth reading not only for its conclusions about phones, which are familiar, but for its proposal that the same methodological errors are about to be repeated, at scale, in the study of generative AI and young people.

The Odgers position should be stated separately because it differs in tone and emphasis from the Orben/Przybylski position. Odgers, in her 2024 Nature review of The Anxious Generation, did not argue that social media is harmless. She argued that the book's central causal claim is not supported by the empirical record, that the convergent population-level evidence is not as strong as Haidt presents it, that other explanations for the post-2012 inflection — including the long-tail effects of the 2008 recession, climate anxiety, the opioid epidemic's effects on family stability, methodological changes in the mental-health survey instruments, and the broader displacement of in-person socialising into mediated forms — are inadequately addressed, and that the policy implications of misattributing the cause matter for the actual adolescent mental health crisis it would be useful to address (Odgers, 2024). Her 2020 Annual Research Review is the longest published synthesis from this side and worth reading in full (Odgers & Jensen, 2020). Its central observation is that the best-designed individual-level studies — preregistered cohort designs, intensive longitudinal methods, ecological momentary assessment — produce small and inconsistent effects, often null, often pointing in different directions depending on platform, use type, and individual vulnerability.

Locating the disagreement

A reader who has stayed with both arguments will notice that the two sides are not, in important respects, claiming opposite things. Most of the apparent contradiction comes from disagreement about what counts as a meaningful effect, not from contradictions in the data themselves.

The Twenge–Haidt side weights population-level changes heavily. The thing that needs explanation, on this account, is that something happened to a generation of teenagers in roughly the same window across most of the developed world, and the timing matches the universalisation of the smartphone and the visual social platform. The exact magnitude of any individual-level correlation is, on this view, almost beside the point: when the population shifts on this many indicators in this short a window, the standard for evidence is different than it is when looking for individual-vulnerability prediction.

The Orben–Przybylski side weights individual-level prediction heavily. The thing that needs explanation, on this account, is that despite decades of effort and large datasets, the best-designed individual-level studies cannot reliably predict which adolescents will be harmed by which uses of which platforms with anything like enough precision to support either a clinical recommendation or a policy intervention. Effects exist, but they are small, conditional, and easily swamped by other factors. The population-level inflection requires explanation, but multiple explanations are compatible with the timing, and the social media one is not strongly favoured by the individual-level data.

These are not the same question. A population-level explanation does not have to predict individual cases. Air pollution is causally implicated in cardiovascular mortality in epidemiological studies; the correlation at the individual level, controlled for confounders, is small. The same pattern is consistent with social media being a real population-level factor that nevertheless has small and noisy individual-level associations. It is also consistent with the population-level inflection having other causes, and the small individual-level associations being mostly confounding. Read carefully, the data does not distinguish cleanly between these two possibilities at the aggregate level.

The most useful empirical move in the past five years has been the attempt to disaggregate — to stop asking does screen time harm teenagers and start asking which uses of which platforms by which subgroups during which developmental windows produce which effects. Several lines of work converge here.

Kelly, Zilanawala, Booker, and Sacker (2019), using the UK Millennium Cohort, looked at 10,904 fourteen-year-olds and found that the association between social media use and depressive symptoms was substantially larger for girls than for boys: a 50% increase in depressive symptom scores for ≥5 hours of daily use in girls, compared to 35% in boys. More importantly, the authors modelled multiple potential intervening routes — online harassment, poor sleep, low self-esteem, and body-image dissatisfaction. The most important routes ran via poor sleep and online harassment; online harassment also linked to depressive symptoms indirectly via sleep, self-esteem, and body image. Because the data are cross-sectional, this should be read as a pathway analysis of associations rather than evidence for direct or indirect causation, and the authors are appropriately careful on that point (Kelly et al., 2019).

Twenge, Haidt, Lozano, and Cummins (2022) re-ran the Orben and Przybylski specification curve analysis on the same underlying datasets, with different constraints: separating boys from girls, separating social media from television and gaming, excluding mediators from the control list, and weighting scales evenly. Among girls, the association between social media use specifically and mental health was -0.11 to -0.24 in median betas — several times larger than the all-screens-and-both-sexes finding of the original analysis, and larger than the corresponding associations in the same data between mental health and binge drinking, sexual assault, or hard-drug use (Twenge et al., 2022). The paper makes a methodological argument: when you aggregate screen time across heterogeneous activities and pool across sexes and developmental stages, you average over the heterogeneity, and the meaningful effects disappear. When you don't, they reappear.

Orben, Przybylski, Blakemore, and Kievit (2022) used UK data on 84,011 participants aged 10 to 80 to examine whether the association between social media use and life satisfaction varied by age. They identified two developmental windows of heightened sensitivity — one centred on early adolescence (around ages 11 to 13 in girls, 14 to 15 in boys) and another around age 19 — in which higher social media use predicted lower life satisfaction the following year, with the effect attenuating outside those windows (Orben et al., 2022). The empirical pattern has a plausible developmental-neuroscience reading in the background: adolescence is a period of heightened peer sensitivity and reward responsivity, which makes the brain unusually receptive to precisely the kind of social feedback the platforms are engineered to deliver. The implication is that the question does social media harm teenagers is too coarse: at minimum, it is does social media harm specific subgroups, during specific developmental windows, when used in specific ways.

If you read these three papers together, the disaggregated picture begins to converge with the Twenge–Haidt position more than the public version of the Orben–Przybylski position suggests. The aggregate effects are small. The disaggregated effects — when you isolate (a) girls, (b) social media specifically rather than screens in general, (c) certain developmental windows, and (d) certain platforms or use types — are several times larger. The clinical relevance of even those larger effects is still debated, but the gap between sides is narrower than the polemics suggest.

The natural experiments

The strongest causal evidence in the field — necessarily — comes from interventions: randomised trials, deactivation experiments, and policy-driven phone bans that create natural experiments at scale. The empirical record here is mixed and instructive.

Melissa Hunt and colleagues at the University of Pennsylvania ran the first sustained-duration RCT explicitly limiting social media across multiple platforms (Hunt, Marx, Lipson, & Young, 2018). One hundred and forty-three undergraduates were assigned either to limit Facebook, Instagram, and Snapchat to ten minutes per platform per day, or to use them as usual. After three weeks, the limited-use group showed significant reductions in loneliness and depression compared to the control group. The effect was largest in participants who had begun the study with higher depression scores. The study is the most-cited proof-of-concept that voluntary, sustained reduction can produce measurable mental health gains; it has been criticised for its undergraduate sample, the absence of an attention-control condition, and the modest absolute effect sizes, but the direction of effect is consistent and the design is clean.

Allcott, Braghieri, Eichmeyer, and Gentzkow (2020) ran a much larger experiment, published in the American Economic Review: 2,743 adult Facebook users were paid to deactivate Facebook for four weeks ahead of the 2018 US midterm elections. Deactivation reduced online activity, reduced political polarization and factual news knowledge in roughly equal measure, increased subjective well-being, and produced a persistent post-experiment reduction in Facebook use. The well-being effects were small in absolute terms but not in proportional terms: subjective well-being rose by about 0.09 standard deviations; Allcott et al. benchmark this as roughly 25–40% of the effect of standard positive-psychology interventions, just over one-third of the conditional wellbeing difference associated with college completion, or roughly the conditional wellbeing difference associated with about USD 30,000 of additional household income (Allcott et al., 2020). The sample was adult, not adolescent, but the design was clean and the effect was on the wellbeing side, in the direction the social-media-harms hypothesis predicts.

The school phone ban literature is newer, more ambiguous, and more important politically. Norway provided the first credible quasi-experimental dataset: Sara Abrahamsson at the Norwegian School of Economics used the staggered timing of phone-ban introduction across Norwegian middle schools to estimate effects on grades, bullying, and mental health (Abrahamsson, 2024). In ban schools, the number of specialist mental-health visits by middle-school girls declined; consultations with general practitioners for mental health complaints declined by roughly 29%; bullying declined for both sexes; girls' grades and externally-marked exam performance improved. Diagnoses of psychological conditions did not change. The Norwegian data is the strongest evidence to date that school-day phone restriction produces measurable population-level benefits, particularly in girls — though it is worth noting that in most of the "ban" schools the policy was less than a total prohibition. Abrahamsson reports that around 45% of ban schools enforced a strict policy; the more common pattern was to allow phones under conditions that did not disrupt class.

A different result has come from the United Kingdom. Goodyear and colleagues ran the SMART Schools study, comparing mental wellbeing, phone use, and screen-time outcomes in 30 secondary schools across England (n = 1,227 adolescents aged 12 to 15) with permissive versus restrictive smartphone policies (Goodyear et al., 2025). The cross-sectional finding was that there was no significant difference in mental wellbeing or out-of-school screen time between restrictive and permissive schools. A 2026 Dutch study compared partial bans (classroom only) with full bans (whole school grounds) across 24 Dutch secondary schools (n = 1,398; Vanluydt et al., 2026) and found, similarly, no significant differences in well-being or bullying outcomes between ban types — and some evidence that full bans were associated with lower student–teacher connectedness and, in girls, reduced school belonging.

A defensible reading of this literature is that bans appear to do something in some places and not in others, and the difference is partly about what counts as a ban (a silent-phone policy is not the same as locked-away phones), partly about the implementation context (a policy nested in a country that prizes school as a phone-free environment will behave differently than one imposed against the grain of local norms), and partly about the outcome measured (mental health service utilisation in Norway behaves differently from cross-sectional self-reported wellbeing in the UK). The most defensible targets are specific time windows — especially the school day and overnight — rather than attempts to remove the device from adolescent life altogether.

The 2024 reckoning

The publication of The Anxious Generation in March 2024 set off the most public phase of the debate so far. The book reached the top of the New York Times bestseller list, drove state-level legislative activity in the United States on school phone policies, and prompted the relevant academic community to respond in unusual depth.

Candice Odgers' review in Nature, published 29 March 2024, set the tone for the academic counter-position. The piece did not contest the existence of an adolescent mental health crisis. It contested the certainty of the diagnosis. The bold claim that digital technologies have rewired children's brains and caused a measurable epidemic, she wrote, is not supported by the science. The hundreds of researchers who have actively looked for the large effects the book asserts have produced a mix of no, small, and mixed associations, most of which is correlational (Odgers, 2024). The danger she identified is not that Haidt is alarmist — alarm may be warranted — but that misdiagnosing the cause distracts from interventions that might actually address the underlying problem.

Several factually grounded criticisms have appeared since the book's release. One is that the cross-national pattern is less uniform than the book implies. PISA school loneliness rose in almost every country measured between 2012 and 2018, but broader well-being and mental health indicators do not move in one clean global direction, and country-level internet-adoption studies find mostly small, null, or mixed associations. A second is that the heavy/light user comparisons in Twenge and Campbell (2019) and similar papers are nearly all cross-sectional, leaving the direction of causality at the heavy-use end plausibly bidirectional: depressed adolescents may use phones more, not only be made more depressed by them.

The policy debate has also moved faster than the evidence. The US Surgeon General's 2023 advisory and the American Psychological Association's 2023 health advisory both framed social media as a plausible youth-risk environment rather than a settled single cause. Australia's under-16 social-media law, passed in November 2024 and in force from 10 December 2025, turned the question into a live regulatory experiment whose outcomes the evidence base will eventually have to reckon with one way or the other.

Vuorre and Przybylski's 2024 paper in Clinical Psychological Science made the population-level case as strongly as it has yet been made on this side of the debate. Using Gallup World Poll wellbeing data from 168 countries between 2005 and 2022, alongside Global Burden of Disease mental-health estimates from 2000 to 2019, the authors examined the association between national-level adoption of the internet and mobile broadband and the corresponding wellbeing and mental-health outcomes. They found small and inconsistent country-level associations, with some countries showing positive effects and others negative (Vuorre & Przybylski, 2024). The result is not a refutation of the Haidt thesis — country-level analyses cannot rule out specific within-population subgroup effects — but it is incompatible with the simplest version of the great-rewiring claim, which would predict broadly consistent negative effects across the countries where internet adoption rose fastest in the relevant window.

The 2024 debate did not resolve. Both sides marshalled additional evidence. Both sides made some methodological concessions. The middle-ground position — that there are real, conditional effects on certain subgroups during certain developmental windows from certain forms of use, that the population-level magnitude is contested, and that the policy interventions worth running are narrower than a generational alarm would suggest — is the position best supported by the disaggregated literature. It is also the position the public discussion has done least to articulate.

What the evidence supports

It is worth distinguishing the confidence levels of the different claims this essay has touched on, because they are not equivalent and the conclusion should not flatten them.

The population-level inflection in adolescent mental health, beginning around 2012 in Anglophone countries: strongly supported. Multiple independent measures (depression diagnoses, self-harm presentations, suicide rates in girls 10 to 14, self-reported loneliness), multiple independent datasets, multiple independent research teams, broadly consistent timing. This is not seriously disputed in the literature (Twenge, 2020; Twenge et al., 2021).

The cross-national consistency of the inflection in roughly the same window: mixed. The PISA loneliness data showing increases in 36 of 37 countries (Twenge et al., 2021) is the most concentrated piece of evidence on this side, but the underlying clinical indicators (depression, self-harm, suicide) are less uniformly available cross-nationally, and the country-level internet-adoption analysis in Vuorre and Przybylski (2024) does not find the consistent negative pattern that the simplest version of the harm hypothesis would predict.

The causal attribution of the inflection to smartphone and social media adoption: weak in the simple form, moderate in the disaggregated form. The simple version — that the rise of the smartphone and the visual platform caused most of the inflection — is consistent with the timing but not strongly favoured by individual-level data, and is in tension with country-level patterns where the speed of internet and smartphone adoption does not predict the size or direction of wellbeing change. The disaggregated version — that certain platforms, used in certain ways, by certain subgroups, during certain developmental windows, account for a substantial fraction of the heavier-use end of the wellbeing decline — is moderately supported (Kelly et al., 2019; Twenge et al., 2022; Orben et al., 2022; Odgers & Jensen, 2020).

The mechanism via specific pathways — online harassment, displaced sleep, appearance comparison on visual platforms, displacement of in-person interaction: moderately to strongly supported, depending on pathway. Online-harassment-to-depression and visual-platform-to-body- dissatisfaction are the strongest pathways in mediator analyses (Kelly et al., 2019; Fardouly et al., 2020). Sleep displacement is well-supported but the magnitude is small (about nine minutes of weekday sleep per additional hour of digital use, per the time-use diary studies); the strength of the downstream sleep–mental-health link is the larger uncertainty (Orben & Przybylski, 2020). The passive- scrolling/rumination pathway has more theoretical than direct empirical support.

The intervention evidence: mixed and conditional. Reducing social media use for a sustained period in motivated participants produces small but consistent wellbeing gains (Hunt et al., 2018; Allcott et al., 2020). School-day phone restriction produces measurable benefits in some implementations (Norway) but not in others (UK, Netherlands), with the difference plausibly traceable to ban design and implementation context (Abrahamsson, 2024; Goodyear et al., 2025; Vanluydt et al., 2026).

The strong policy claims — blanket bans, age limits, platform-level restrictions justified by individual-level mental health benefit: weak to moderate. The evidence base is strong enough to support specific, narrow interventions (school-day restriction, overnight phone separation) but not strong enough to support the sweeping claim that the modal effect on the modal teenager from the modal use is harmful. Age-based restrictions on visual platforms for early adolescents are a defensible direction given the disaggregated findings, but they are a stronger policy claim than the evidence has yet directly tested.

Reading the literature this way leaves the reader with less clarity than the popular discussion offers, which is itself the point. The evidence supports a more local and conditional set of actions than either Haidt's book or the strongest critics of it would frame as urgent. The evidence does not support inaction.

For an individual reader

If you are reading this as someone trying to figure out what to do — for yourself, for a teenager you are responsible for, or simply as someone trying to understand what is and isn't known — the literature supports the following set of suggestions, in roughly descending order of evidentiary strength.

1. Protect sleep. The evidence linking phone-in-the-bedroom to displaced sleep is consistent across the methodologically tighter studies, and the evidence linking adolescent sleep deprivation to mental health is older and stronger still. Phones outside the bedroom overnight is the simplest, best-supported intervention. (Orben & Przybylski, 2020)

2. Treat appearance-centred visual platforms as a higher-risk category, especially for early-adolescent girls. The most consistent sub-group findings in the disaggregated literature involve appearance comparison on visual platforms in early-adolescent girls. Instagram, Snapchat, and TikTok are not interchangeable with messaging in their effects. The Kelly et al. (2019) pathway analysis points specifically at body image and online harassment as routes from social media use to depressive symptoms; the Twenge et al. (2022) specification-curve reanalysis shows that the girls-and-social-media interaction is substantially larger than the all-screens-and-both-sexes finding. Delaying or restricting these platforms where practical is reasonable. (Fardouly et al., 2020)

3. Restrict school-day use. The Norwegian quasi-experimental data is the strongest population-level evidence the field has on the mental health effects of phone restriction. The effect was concentrated in girls. The most plausible mechanism is not mysterious; it is that displaced peer interaction during the school day reduces opportunities for the social-comparison, harassment, and rumination loops the within-school day otherwise mediates. (Abrahamsson, 2024)

4. Distinguish use type and online experience. The literature is increasingly clear that time on device is too blunt. Messaging a friend, being targeted by online harassment, compulsively checking an app, scrolling an appearance-centred feed, and using Reddit or dating apps are not interchangeable exposures. Matthews et al. (2025) found WhatsApp use associated with lower loneliness, while Reddit and dating-app use, compulsive use, and online victimisation were associated with higher loneliness. That is the level of specificity the practical advice should preserve. (Matthews et al., 2025; Orben et al., 2022)

5. Hold the policy claims more loosely than the personal ones. The evidence supports narrower interventions than the public discussion is currently structured around. A blanket ban on smartphones for under-16s is a strong policy claim whose evidence base is more contested than the personal-use claims in (1) to (4). Whether the right level of action is the individual, the family, the school, the platform, or the regulator depends on which intervention is being considered, and the literature does not deliver clean answers at the policy scale.

The shortest defensible summary is unromantic. There is a real, measurable, post-2012 adolescent mental health worsening in the developed world. There is a real, conditional, sub-group-specific contribution from social media and smartphones to that worsening, particularly for girls and particularly through visual platforms, online harassment, and displaced sleep. There is a much larger gap between what the data show and what the loudest commentators on each side claim than the public discourse permits. And there is, in the carefully disaggregated literature of the past five years, a converging picture that supports specific, narrow, evidence-based interventions in particular contexts — and counsels against the kind of confident generational diagnosis that has dominated the popular conversation.

Adolescents have been growing up with these devices for fifteen years now. They are not, on the available evidence, a generation broken by their phones. They are also not, on the available evidence, unaffected. The right reading is the third one, which is also the boring one: there are real harms in real places to real subgroups, and there are useful things that can be done about them. Most of those useful things are smaller and more specific than the alarm suggests. That is, on the evidence, the responsible reading.

---

References

• Abrahamsson, S. (2024). Smartphone Bans, Student Outcomes and Mental Health. NHH Department of Economics Discussion Paper 01/2024, Norwegian School of Economics.
• Allcott, H., Braghieri, L., Eichmeyer, S., & Gentzkow, M. (2020). The welfare effects of social media. American Economic Review, 110(3), 629–676.
• Fardouly, J., Magson, N. R., Rapee, R. M., Johnco, C. J., & Oar, E. L. (2020). The use of social media by Australian preadolescents and its links with mental health. Journal of Clinical Psychology, 76(7), 1304–1326.
• Goodyear, V. A., Randhawa, A., Adab, P., Al-Janabi, H., Fenton, S., Jones, K., Michail, M., Morrison, B., Patterson, P., Quinlan, J., Sitch, A., Twardochleb, R., Wade, M., & Pallan, M. (2025). School phone policies and their association with mental wellbeing, phone use, and social media use (SMART Schools): a cross-sectional observational study. The Lancet Regional Health – Europe, 51, 101211.
• Haidt, J. (2024). The Anxious Generation: How the Great Rewiring of Childhood Is Causing an Epidemic of Mental Illness. Penguin Press.
• Hunt, M. G., Marx, R., Lipson, C., & Young, J. (2018). No More FOMO: Limiting Social Media Decreases Loneliness and Depression. Journal of Social and Clinical Psychology, 37(10), 751–768.
• Kelly, Y., Zilanawala, A., Booker, C., & Sacker, A. (2019). Social media use and adolescent mental health: Findings from the UK Millennium Cohort Study. EClinicalMedicine, 6, 59–68.
• Mansfield, K. L., Ghai, S., Hakman, T., Ballou, N., Vuorre, M., & Przybylski, A. K. (2025). From social media to artificial intelligence: improving research on digital harms in youth. The Lancet Child & Adolescent Health, 9(3), 194–204.
• Matthews, T., Arseneault, L., Bryan, B. T., Fisher, H. L., Gray, R., Henchy, J., Moffitt, T. E., & Odgers, C. L. (2025). Social media use, online experiences, and loneliness among young adults: A cohort study. Annals of the New York Academy of Sciences, 1548(1), 194–205.
• Odgers, C. L. (2018). Smartphones are bad for some teens, not all. Nature, 554(7693), 432–434.
• Odgers, C. L. (2024). The great rewiring: is social media really behind an epidemic of teenage mental illness? Nature, 628, 29–30.
• Odgers, C. L., & Jensen, M. R. (2020). Annual Research Review: Adolescent mental health in the digital age: facts, fears, and future directions. Journal of Child Psychology and Psychiatry, 61(3), 336–348.
• Orben, A., & Przybylski, A. K. (2019a). The association between adolescent well-being and digital technology use. Nature Human Behaviour, 3(2), 173–182.
• Orben, A., & Przybylski, A. K. (2019b). Screens, teens, and psychological well-being: Evidence from three time-use-diary studies. Psychological Science, 30(5), 682–696.
• Orben, A., & Przybylski, A. K. (2020). Teenage sleep and technology engagement across the week. PeerJ, 8, e8427.
• Orben, A., Przybylski, A. K., Blakemore, S.-J., & Kievit, R. A. (2022). Windows of developmental sensitivity to social media. Nature Communications, 13, 1649.
• Twenge, J. M. (2020). Increases in Depression, Self-Harm, and Suicide Among U.S. Adolescents After 2012 and Links to Technology Use: Possible Mechanisms. Psychiatric Research and Clinical Practice, 2(1), 19–25.
• Twenge, J. M., & Campbell, W. K. (2019). Media Use Is Linked to Lower Psychological Well-Being: Evidence from Three Datasets. Psychiatric Quarterly, 90(2), 311–331.
• Twenge, J. M., Haidt, J., Blake, A. B., McAllister, C., Lemon, H., & Le Roy, A. (2021). Worldwide increases in adolescent loneliness. Journal of Adolescence, 93, 257–269.
• Twenge, J. M., Haidt, J., Lozano, J., & Cummins, K. M. (2022). Specification curve analysis shows that social media use is linked to poor mental health, especially among girls. Acta Psychologica, 224, 103512.
• Vanluydt, E., van den Eijnden, R., Vonk, L., Putrik, P., van Amelsvoort, T., Delespaul, P., Levels, M., & Huijts, T. (2026). Disconnect To Reconnect: How Variations between Types of Smartphone Bans Influence Students' Well-being and Social Connectedness in Dutch Secondary Education. Journal of Youth and Adolescence, 55(3), 551–569.
• Vuorre, M., & Przybylski, A. K. (2023). Estimating the association between Facebook adoption and well-being in 72 countries. Royal Society Open Science, 10(8), 221451.
• Vuorre, M., & Przybylski, A. K. (2024). Global well-being and mental health in the internet age. Clinical Psychological Science, 12(5), 917–935.