The Anxious Generation is our book club pick this month, which is the best reason I can think of to have read it carefully. Book clubs are good at exactly this. You end up with a book you would have skimmed or skipped, and a room full of people who are going to ask you what you think, so you have to go find out. Half the value is the friction. You read something you half agree with, someone else read it differently, and you leave having tested a position instead of just holding one.
So I went and read the underlying research, both the case against the book and the case for it. What I found is that the argument everyone is having is not the argument the evidence is having.
Start here. In one of the most cited studies in this whole fight, wearing glasses was more strongly associated with teenage unhappiness than screen time was. Eating potatoes was almost as bad as both. That comes from Amy Orben and Andrew Przybylski's 2019 paper in Nature Human Behaviour, which ran the numbers across 355,358 adolescents and found that digital technology use explained at most about 0.4 percent of the variation in well-being.
The potato line gets quoted like a mic drop. It isn't one, and understanding why it isn't is the fastest way into the real disagreement about Jonathan Haidt's book.
Here is what the line does establish. Ordinary hours of ordinary screen use, measured by asking teenagers how much they use screens, has a tiny person-level relationship with how they say they feel. That finding has held up. Fassi and colleagues, in JAMA Pediatrics in 2024, pooled 143 studies covering 1,094,890 adolescents and found correlations around r = .08 for time spent and r = .12 for engagement measures. Small.
Here is what the line does not establish. It says nothing about what happens when an entire cohort gets a technology at once. A risk factor with a small average effect and 95 percent exposure can move a population distribution more than a large effect that touches 3 percent of people. That is not a rhetorical dodge; it is how epidemiology works. It is also the exact argument the tobacco industry lost.
So the fight is not about whether the correlations are small. Both sides agree they are small. The fight is about whether small correlations measured badly can add up to a real population shift, and nobody in this literature has produced the number that would settle it.
That number has a name. It is the population attributable fraction, the share of a trend that would disappear if you removed the exposure. Of the total rise in teenage depression and suicide since 2010, how much of it goes away in a world without smartphones? Sixty percent? Five? Nobody has computed it, and the reason is that computing it requires four inputs the field does not have: a credible causal effect per unit of exposure, the real distribution of that exposure across ages and platforms, the size of the genuinely susceptible group, and what the exposure displaced. Every argument in this book, and every argument against it, is a guess about that one number.
The two reports agree more than either admits
I read two long research syntheses on the book, one written to make the strongest case against it and one written to make the strongest case for it. They cite many of the same papers. Strip the framing and the overlap is uncomfortable for anyone who wants a clean answer.
Both agree something real happened to teenagers. This is where the pure skeptics overreach. The measurement-artifact story, that changed diagnostic practice and reduced stigma and expanded screening manufactured a fake crisis, explains part of the rise in self-reported sadness. It cannot explain US suicide mortality for ages 10 to 24 rising 62 percent between 2007 and 2021, or the rate among 10 to 14 year olds roughly tripling between 2007 and 2018. It also cannot explain England, where the NHS ran comparable psychiatric assessments with clinically trained raters in 1999, 2004, and 2017, and emotional disorders among 5 to 15 year olds went from 4.3 percent to 3.9 percent to 5.8 percent.
Both agree the causal evidence is no longer purely correlational. The single strongest study in the debate is Braghieri, Levy, and Makarin in the American Economic Review in 2022, which exploited the staggered rollout of Facebook across US colleges. Students did not choose when Facebook arrived at their school. Mental health got worse by 0.077 standard deviations after it did, implying roughly a 9 percent relative rise in predicted depression and 12 percent in anxiety. Then there are the randomized trials. A 2025 meta-analysis of 10 randomized controlled trials with 1,491 participants found that cutting social media use reduced depressive symptoms by Hedges' g = .25 after correcting for publication bias. Anyone still saying "it's only correlation" stopped reading in 2019. Confidence is high that there is a real causal effect somewhere in here.
Both sides agree the effect is not close to established as the dominant cause. Even the supportive report lands on moderate confidence for "major population contribution" and moderate-low for "dominance."
The disagreement, in other words, is entirely about magnitude. And magnitude is the one thing the field has not measured.
The anomaly worth chasing
Start with dates, because dates are cheap to check and hard to spin.
US youth suicide starts climbing in 2007. Smartphone ownership among American teenagers was 23 percent in 2011. It hit 73 percent around 2014 to 2015 and 95 percent by 2018. Haidt's "Great Rewiring" is dated 2010 to 2015.
The hard outcome moves first. By several years. That is a genuine problem for a story where a 2012 technology shock causes a mental health collapse, and the supportive report concedes it directly rather than hiding it, which is to its credit. The obvious rival explanation sitting in that window is the financial crisis and what it did to families.
Defenders answer that Facebook and broadband and feature phones were already spreading, and that different outcomes need not break in the same year. Fair enough. But notice what that answer costs. It converts a sharp, dateable, falsifiable claim about smartphones into a fuzzy claim about digital life generally, which is much harder to test and much easier to defend from anything.
The same softening happens with the best evidence. The Facebook rollout study is the strongest causal finding Haidt has. It is about college students in the mid 2000s, on a desktop social network, before the iPhone existed. It demonstrates that online social comparison can hurt mental health. It does not demonstrate that the phone-based childhood of an eleven year old is the mechanism at work. The most persuasive support for the book comes from a period the book treats as the before picture.
Where the evidence gets genuinely contradictory
Two studies published within a year of each other, using the same statistical method designed to separate within-person change from between-person differences:
The Adolescent Brain Cognitive Development cohort, 11,876 children followed over four waves, found that when a child's social media use rose above their own usual level, depressive symptoms rose the following year, with standardized coefficients of .07 and .09. The reverse path was absent. Depression did not predict later social media use.
A 2026 study of 25,629 adolescents in Greater Manchester, three waves, same modeling approach, found nothing. No prospective effect on internalizing symptoms in boys or girls, and splitting active from passive use did not rescue it.
That is not a case of good studies versus bad studies. Those are both the design that critics kept demanding, run on large samples, producing opposite answers. When your best methods disagree, the honest reading is that the effect is small enough to be swamped by sample and measurement differences, and heterogeneous enough that population averages are the wrong summary statistic.
The neuroscience is cleaner and worse for Haidt. The book's central metaphor is rewiring. Miller, Mills, Vuorre, Orben, and Przybylski looked at over 4,000 children in the ABCD imaging data and found that screen media activity was not meaningfully related to functional brain organization, and heavier users did not show maladaptive profiles. Neither preregistered hypothesis survived. The largest study of American adolescent brain development finds no rewiring. Confidence in other words is high that the neural claim is unsupported. Haidt is using a mechanism word where he has an epidemiology argument.
The aggregation problem is the real bug
Here is the thing that bothers me most, and it is a design flaw rather than a data problem.
"Social media" is not one treatment. WhatsApp with three close friends, Discord with a gaming group, YouTube tutorials, and a 1 a.m. algorithmic feed of appearance-ideal video are different experiences that share a delivery device. Studying them together is like studying "video games" by averaging Tetris and Call of Duty and concluding something about screens. The average is real arithmetic and describes no actual player.
Watch what happens when researchers stop aggregating. De Valle and colleagues ran an experimental meta-analysis on appearance-ideal social media content and found it worsened body image with Hedges' g = -.61, a moderate effect and roughly six times the size of the generic time-spent correlation. John and colleagues, across 26 independent studies and 156,384 young people, found cyber-victims had odds ratios of 2.35 for self-harm and 2.57 for suicide attempts. A 2026 randomized trial in a distressed youth sample, where 75 percent exceeded the depression threshold at baseline, found that cutting social media to one hour a day produced a clinically meaningful anxiety improvement in the high-anxiety group specifically.
The signal is not evenly spread across the exposure. It concentrates in particular content, particular kids, particular hours of the night. And the mechanisms Haidt leans on hardest do not survive equally. Sleep, which sounded like his most physiological argument, weakened badly under better methods: a 2026 JAMA Pediatrics meta-analysis of within-person daily data found screen time predicted slightly later sleep onset at r = .079 and no significant effect on total sleep time or quality. Attention fragmentation has good laboratory evidence for acute task disruption and almost no bridge to psychiatric outcomes.
Haidt's four foundational harms are not four equally supported pillars. Two are decent, one is weaker than advertised, and one is mostly assertion.
The half he is most confident about is the half he can least support
The book has two arguments. Childhood lost its unsupervised play and independence, starting in the 1980s. Then phones arrived and finished the job. Haidt presents these as a matched pair.
They are not matched at all in evidentiary status. The phone half is testable and has been tested hard, with natural experiments, large cohorts, meta-analyses, and randomized trials. The play half rests on developmental theory, a three-wave study of 471 children showing bidirectional effects, a preregistered Dutch study of 2,229 adolescents establishing concurrent associations, and a 2024 systematic review of 23 studies on independent mobility that concluded the psychosocial evidence remains sparse. Confidence is low to moderate that lost independence is a quantitatively major cause of the post-2010 inflection.
This creates a bundling problem that I think is the book's most consequential rhetorical move, and it cuts against Haidt in a way his critics have underused. He pairs a claim most parents already believe and that costs nothing to accept, that kids need more freedom and less supervision, with a claim that carries enormous policy weight, that platforms caused a mental illness epidemic. The first lends credibility to the second. Neither has to earn it separately.
And if you take the overprotection argument seriously, it becomes a rival to the phone argument rather than a partner to it. Adults spent thirty years removing autonomy, unstructured risk, and unsupervised peer time from childhood. Then those same adults gave kids a device and blamed the device. It is at least possible that the phone is where displaced developmental needs went rather than what displaced them.
What the interventions show, which is the part that should decide it
If phones are a major cause, removing them should do something measurable. This is the closest thing to a live experiment we have, and it is not going Haidt's way so far.
The SMART Schools study, published in Lancet Regional Health Europe in February 2025, compared 1,227 students across 30 English schools, 20 with restrictive phone policies. No difference in mental wellbeing, anxiety, depression, sleep, physical activity, or attainment. The restrictive schools cut in-school phone use by about 40 minutes and social media by about 30 minutes, and did not reduce overall daily use meaningfully. Confidence is high on the finding, moderate on its generalizability given the cross-sectional design.
Sara Abrahamsson's work on roughly 400 Norwegian middle schools found no average effect but real gains for girls, with psychological consultations down about 29 percent and specialist visits down close to 60 percent. Early US evidence is mixed, including an NBER analysis of state bans finding no clear mental health improvement and a pouch study finding well-being dropped before it rose.
Australia went furthest, banning under-16s from major platforms from December 2025. Early reporting suggests teenagers are routing around it, which is what the digital rights researchers and the child development researchers both predicted, for different reasons.
A theory that predicts large harm from an exposure should predict recovery when you remove the exposure. The removal experiments are producing small, girl-specific, implementation-dependent effects. That is consistent with a modest real effect concentrated in a susceptible subgroup. It is not consistent with an epidemic driven by a single environmental cause.
The verdict
Haidt is directionally right and quantitatively overclaiming, and the overclaim is doing more damage than the underlying error.
Confidence is high that youth mental health genuinely deteriorated. Confidence is high that social media is a causal contributor to some of it. Confidence is moderate that it is a meaningful population-level contributor. Confidence is low that it is the dominant cause. Confidence is high that the brain-rewiring framing is unsupported. Confidence is low to moderate that the lost-play half explains the 2010s inflection at all.
The missing piece is a population attributable fraction, the estimate of what share of the trend the exposure explains, built from credible causal effects across the real distribution of ages, sexes, platforms, doses, and displaced activities. Braghieri and colleagues attempted the only serious version of it and got roughly 24 percent of the rise in severe depression among college students, while flagging heavy assumptions. One tentative number, from the wrong population, in the wrong decade. That is what the strongest claim in this debate rests on.
Until someone produces that number, the correct posture is to separate the policies by cost.
Cheap and reversible: phones out of classrooms, which is defensible on attention and learning grounds without needing the mental health argument at all; later first-phone ownership, which costs a family nothing to try; more independence and unsupervised time, which is good developmental practice regardless of who is right about Instagram. Do all of that now. The evidence does not have to be conclusive to justify actions this cheap.
Expensive and hard to unwind: national age verification, government identity infrastructure, blanket platform bans. These require identity data for every user, create breach surfaces, and fall hardest on the teenagers whose online communities substitute for a hostile local one. They should require the number, and the number does not exist.
The most interesting thing about this whole episode is not the science. It is that Haidt ran a public advocacy campaign in parallel with the scientific argument and largely won the public one before the empirical one resolved, which changed what the evidence was even for. Whether that is admirable urgency or motivated reasoning depends on whether he turns out to be right, and we will not know for years. But it is a repeatable playbook, and someone will run it again on a weaker case. The lesson worth carrying is smaller and more useful than the book's thesis. When a claim arrives pre-bundled with a policy agenda and a movement, go find the number that would have to be true, and check whether anyone has it.