The share of U.S. adolescents who had a major depressive episode in the past year was roughly flat from 2004 to 2011. Then it began rising. It hasn't stopped.
Past-year major depressive episode, U.S. adolescents ages 12 to 17. Source: SAMHSA National Survey on Drug Use and Health (NSDUH), 2004 to 2021. NSDUH was redesigned in 2020. Pre- and post-2020 figures are not strictly comparable. The 2011 inflection shows up across surveys and across countries either way.
It is not only the United States.
The U.K., Canada, Australia, New Zealand, and the Nordic countries show the same inflection. Same direction. Same timing. Same magnitude. Every English-speaking democracy that adopted smartphone-based adolescence in that window saw teenage mental health turn at the same point.
Six countries do not turn in lockstep by accident. Something common to all of them changed in those years. There is one obvious candidate.
A correlation.
A causal claim.
Many things rose in the 2010s. Smartphone adoption. Algorithmic feeds. Sleep loss. Helicopter parenting. Political polarization. Climate anxiety. A striking chart is not the same as proof.
The strongest causal evidence in this literature is narrower and quieter than the chart above.
01Allcott, Braghieri, Eichmeyer, and Gentzkow (2020). Randomly paid 2,743 U.S. adults to deactivate Facebook for four weeks. Measured improvements in subjective well-being. A real experiment, not a correlation.
02Haugen (2021), the Facebook Files. Meta's own internal research, leaked: Instagram knew it made body image worse for one in three teen girls. Internal causal language. Internal data.
03Twenge & Haidt (2024). Cross-country replication: the same generational inflection across the English-speaking world plus the Nordics. Coordinated timing rules out most local explanations.
The question we cannot run.
What would the curve above look like if smartphone-based adolescence had not happened? If Instagram had shipped with the age gates Meta's own researchers recommended? If TikTok's recommender weighted differently? If schools had banned phones in 2014?
These are counterfactuals. You cannot randomize a generation. The human-subjects experiment that would answer the question is the one we are not allowed to run. The one already running, on every adolescent alive, is the one we cannot undo.
Subconscious builds AI agents that simulate populations under conditions that did not occur. Counterfactual evidence at the scale of real markets. The same machinery that lets a brand test a campaign before launch lets a policymaker test an intervention before passing it. We're applying it to the questions human-subjects research will never run.
Every cost estimate in this debate hides five judgment calls. Drag the sliders. Watch the number change. You're now the one making the claim.
Annual cost to the United States
$0
A point estimate from your assumptions. Not a forecast. Not a verdict. A way to see which judgment calls do the work.
Mortality
$0
Mental health
$0
Economic
$0
How much of the rise do we blame on social media?
18%
Source
Twenge & Haidt 2024 argue the share is large. Skeptics like Odgers and Orben argue under 10%. The debate is the point.
How many U.S. teens and young adults have depression today?
5.0M
Source
SAMHSA NSDUH 2021 implies roughly 5M. Survey-based and clinical estimates diverge by a factor of two.
What's a statistical life worth?
$13.7M
Source
DOT 2024 guidance sets VSL at $13.7M. EPA uses roughly $11M. Academic estimates span $5M to $20M.
How many years does it last per person?
6 yrs
Source
Onset-to-recovery in clinical samples runs shorter. Longitudinal cohorts run longer. Both are real.
How long do treatment costs run?
4.5 yrs
Source
Claims-data studies suggest a few years of elevated cost. The identification problem is the same as on the mortality side.
Default total. About $2.5T.
Fixed assumptions. Annual U.S. suicides 110,000 (CDC WONDER, all ages, recent five-year mean). Quality-of-life loss per depression-year 35% (Salomon et al. 2012, GBD disability weights). Healthcare cost $7,000 per person per year and productivity loss $6,000 per person per year (Greenberg et al. 2021, economic burden of major depressive disorder in the U.S.). These are accounting parameters, not causal claims.
These are the numbers the world is arguing about. Subconscious builds the agents that let you stress-test them. subconscious.ai