The study where nobody could tell

One hundred and forty-four Dutch girls, aged 14 to 18, were shown Instagram selfies in a lab. Half saw the original photographs. Half saw versions in which the bodies had been reshaped. Then they were asked how realistic the images were.

They could not tell. Both groups rated the pictures as realistic. The girls who had seen the manipulated bodies rated those images more highly than the control group rated the originals. And their own body image was worse afterwards.

That is Kleemans, Daalmans, Carbaat and Anschütz, “Picture Perfect: The Direct Effect of Manipulated Instagram Photos on Body Image in Adolescent Girls”, published in Media Psychology 21(1): 93–110 in 2018. It is a between-subjects experiment with random assignment — the design that licenses causal language. Exposure to manipulated photos, the authors report, directly led to lower body image, and the effect was strongest among girls with a higher tendency to compare themselves to others.

The finding that matters most is the one about detection. A defence people offer for the whole retouching economy is that everyone knows. Everyone can see it. Nobody is fooled. In this manipulated Instagram photos body image study, nobody could see it, and the undetected version worked better than the truth.

Why the strongest evidence is not the loudest

Here is where most writing on this subject goes wrong, and where this publication is going to be tediously careful instead.

The literature on social media and adolescent mental health is genuinely contested. Effect sizes in correlational work are typically small. Prominent researchers disagree sharply — not about the data, which is largely shared, but about whether it supports causal claims at all. If you have read a confident article saying the science is settled in either direction, you have read someone flattening a live argument.

The reason for the disagreement is structural. Most of the evidence is correlational: survey a lot of teenagers, ask how much they use Instagram, ask how they feel, look for a relationship. When you find one, you cannot tell which way it runs. Unhappy adolescents may use more social media because they are unhappy. Something else — sleep, family, school, poverty — may be driving both. And the associations that emerge from these designs are usually small enough that ordinary analytic choices can move them around.

So rank the evidence by design, not by headline. Two studies deserve most of the weight.

The staggered rollout

Luca Braghieri, Ro’ee Levy and Alexey Makarin published “Social Media and Mental Health” in the American Economic Review 112(11): 3660–3693 in 2022. It is the most methodologically rigorous causal evidence available on this question.

Facebook launched college by college, in an order driven by the company’s own expansion plans rather than by anything about the students. That makes the rollout a natural experiment: some campuses got Facebook earlier than others for reasons unrelated to their mental health. The authors exploit this with a difference-in-differences design, comparing changes in student mental health at campuses that had just received Facebook against changes at campuses that had not yet.

The finding: Facebook’s arrival worsened student mental health and increased reports that poor mental health had impaired academic performance. The mechanism analysis points to unfavourable social comparisons as the driver.

This design answers the objection that sinks the survey work. Students did not choose to have Facebook arrive; it arrived. Reverse causation is ruled out by the timing, and unmeasured differences between campuses are absorbed by the comparison structure.

Its limits are real and should be stated. It studies Facebook in 2004–2006, a text-and-profile product very unlike an algorithmic image feed. It studies American undergraduates, not adolescents. And it measures a shock — the arrival of a new platform — rather than the steady state of having grown up inside one.

The ten-minute experiment

Hunt, Marx, Lipson and Young, “No More FOMO: Limiting Social Media Decreases Loneliness and Depression”, Journal of Social and Clinical Psychology 37(10): 751–768, 2018. One hundred and forty-three University of Pennsylvania undergraduates were randomly assigned either to limit Facebook, Instagram and Snapchat to ten minutes per platform per day, or to use them normally, for three weeks. The limited group showed significant reductions in loneliness and depression relative to the control group.

Randomised, so causal. Small, so imprecise: 143 students at one elite university, over three weeks, on self-reported outcomes, with no blinding — participants obviously knew which group they were in, and expectations can move self-report. Treat it as a demonstration that the direction of effect is plausible and reversible, not as an estimate of how large it is.

The earlier experimental work

Fardouly, Diedrichs, Vartanian and Halliwell, “Social comparisons on social media: The impact of Facebook on young women’s body image concerns and mood”, Body Image 13: 38–45, 2015, established an early experimental link between social media exposure and appearance comparison in young women. It is part of the reason Kleemans et al. designed their study around comparison tendency rather than exposure alone.

What the Surgeon General actually said

In May 2023 the US Surgeon General issued an advisory, “Social Media and Youth Mental Health”. A synthesis of 20 studies found significant associations between social media use and body-image concerns, with social comparison as a contributing factor. The advisory’s own formulation is carefully hedged: “Social comparison driven by social media is associated with body dissatisfaction, disordered eating, and depressive symptoms.”

Associated with. Not causes. An advisory is a public-health judgement made under uncertainty, which is a legitimate thing for a public-health official to do — you do not wait for perfect evidence before warning people about a plausible risk. But it is not the same act as establishing a fact, and it should not be quoted as though it were.

The advisory also carries a number worth keeping: 46 per cent of teens aged 13 to 17 said social media makes them feel worse about their body image. Forty per cent said neither better nor worse. Only 14 per cent said better. That is self-report, with all the caveats from the paragraph above about people narrating their own causation. It is still a striking distribution, because of the shape: the option “makes me feel better” is the one almost nobody chooses.

The narrow claim, which is enough

Put the defensible findings together and you do not get “social media causes depression”. You get something smaller and much harder to dismiss.

Exposure to digitally reshaped bodies lowers the body image of adolescent girls in the short run, and they cannot tell which photographs have been reshaped (Kleemans, n=144, randomised). The arrival of a platform built around peer comparison worsened mental health at population scale, and comparison appears to be the mechanism (Braghieri et al., quasi-experimental, AER). Deliberately cutting use reduces loneliness and depression over three weeks (Hunt et al., n=143, randomised).

None of that requires the strong claim. It does not need social media to be uniquely harmful, or harmful to everyone, or the dominant cause of anything. It needs only that unfavourable comparison against images that are better than reality does measurable damage — and that the images are better than reality by design, at industrial volume, in a way the audience cannot see.

That last clause is what the Dutch experiment nailed shut. The most consequential result in this whole literature is not an effect size. It is that 144 teenagers looked carefully at photographs of altered bodies, were asked directly whether the images were realistic, and said yes.