Social media effects are real but smaller, more conditional and harder to measure than the public debate assumes

The public argument about social media is conducted as if the effects were established and only the remedy in dispute. The ninth of ten units finds that the literature supports a narrower claim: the effects are real, usually small, and dependent on who is measured and how. Design explains the disagreement between studies better than any underlying effect does (Vosoughi, Roy and Aral, 2018; Guess, Nyhan and Reifler, 2020b). The tenth unit asks what kind of evidence such effects rest on.

The size of an effect is decided by the design

Guess, Nyhan and Reifler (2020b) combined survey responses with web-traffic records from the 2016 US presidential campaign. Untrustworthy websites were a small share of information diets on average, consumed largely by a subset with strong pro-attitudinal preferences, so speculation about prevalence has been overstated.

Allcott and Gentzkow (2017) reached a compatible estimate from browsing data, fact-checking archives and a survey: fourteen per cent named social media their most important source of election news, the average adult saw one or a few false stories, and just over half who recalled them believed them. The scale depends on whether the analyst measures what was published, what was available or what was consumed, and consumption is the smallest.

Grinberg and colleagues (2019) matched registered voters' Twitter accounts to fake news sources in the 2016 campaign, and found engagement extremely concentrated. Just 1 per cent of individuals accounted for 80 per cent of exposure to fake news sources, and 0.1 per cent for nearly 80 per cent of the sources shared, while most political news still came from mainstream outlets. The same corpus that yields a small average effect yields a severe tail, so the size of the effect and the size of the population it falls on are different questions.

Falsehood diffuses further, and the feed decides less than the person

Vosoughi, Roy and Aral (2018) analysed roughly 126,000 stories tweeted by about 3 million people more than 4.5 million times between 2006 and 2017, classified by six independent fact-checking organisations that agreed on 95 to 98 per cent of the classifications. Falsehood diffused farther, faster, deeper and more broadly than truth in every category, and robots spread both at the same rate, so the mechanism was human sharing.

A diffusion advantage across millions of tweets is compatible with few people seeing any given false story (Allcott and Gentzkow, 2017).

The strongest evidence against a simple algorithmic explanation comes from Bakshy, Messing and Adamic (2015), who used data from 10.1 million US Facebook users to separate what friends shared, what the ranking displayed and what users clicked. Individual choice limited exposure to cross-cutting content more than the algorithm did.

That bounds the filter-bubble claim rather than settling it: it measures one feed in one period and cannot say what the same users would have seen without ranking.

Contact with the other side can inflame polarisation

The contact hypothesis predicts that exposure to opposing views softens them, and the strongest test points the other way. Bail and colleagues (2018) paid a random group of frequent Twitter users to follow a bot posting opposing political views for one month. Republicans who followed a liberal bot became substantially more conservative; the Democratic shift was slight and not significant.

Boxell, Gentzkow and Shapiro (2017) built a polarisation index from eight measures and found polarisation growing fastest among the groups least likely to use the internet, with larger increases among people over 65 than among those aged 18 to 39. The age group measured is part of the design rather than a footnote to it. Each answer is limited by a counterfactual it cannot observe.

The wellbeing result turns on the counterfactual, and participation barely moves

Two strong designs reach opposite conclusions about wellbeing. Allcott, Braghieri, Eichmeyer and Gentzkow (2020) randomised participants to deactivate Facebook for the four weeks before the 2018 US midterm election. Deactivation reduced online activity and polarisation, reduced factual news knowledge and increased subjective wellbeing.

Braghieri, Levy and Makarin (2022) exploited the staggered arrival of Facebook at US colleges and found the rollout worsened student mental health and increased reported academic impairment, with evidence pointing to unfavourable social comparison. Four weeks of absence against years of exposure: the disagreement is about the counterfactual, not arithmetic.

The pooled evidence now speaks to the same question. Ferguson, Kaye, Branley-Bell and Markey (2025) meta-analysed 46 studies of youth social media use and mental health and found the pool unable to support claims of harm, while recording common weaknesses such as demand characteristics and absent preregistration. The estimate is small and the designs weak enough that confident public claims are unsupported.

Boulianne (2015) meta-analysed 36 studies. More than 80 per cent of coefficients were positive, but only half were statistically significant. Panel studies were less likely to report significant effects than cross-sectional surveys, and the campaign impact was minimal.

The mechanism is a failure of reasoning more than a bias

The mechanism claim is well evidenced. Pennycook and Rand (2019) tested more than 3,400 participants on the Cognitive Reflection Test and on the perceived accuracy of fake and real headlines. Higher reflection predicted better discernment even for ideologically aligned headlines, so susceptibility tracked analytic engagement, not partisan motivation.

Pennycook and colleagues (2020) asked more than 1,700 adults to judge the accuracy of an unrelated headline before deciding what to share, and the reminder nearly tripled truth discernment in later sharing. A prompt of a few seconds beat most direct corrections.

The most reliable effect is an incentive to polarise

The largest and most consistent effect in this unit is not on users' attitudes but on the incentive structure around them. Rathje, Van Bavel and Van der Linden (2021) analysed posts from news outlets and political leaders and found out-group animosity the strongest predictor of engagement: each out-group term raised the odds of a share by 67 per cent, about 4.8 times the effect of negative affect.

Whatever the size of an effect on an individual, the platform distributes the content that earns engagement, and anger at the other side earns the most.

Contested question: can better measurement settle the disagreement, or does it turn on which counterfactual is right?

Three positions are on the table. The first holds that the effect is small and the alarm overstated: exposure to untrustworthy sources is concentrated in a minority (Guess et al., 2020b), consumption of false news ran to one or a few stories (Allcott and Gentzkow, 2017), the user's choice limits cross-cutting exposure more than the algorithm (Bakshy et al., 2015), and the groups polarising fastest have the least internet use (Boxell et al., 2017).

The second holds that the effects are real and detectable only under the right counterfactual, so the small estimates are artefacts of weak designs. Braghieri, Levy and Makarin (2022) find mental-health harm where a four-week deactivation finds improved wellbeing (Allcott et al., 2020).

The third locates the durable effect in the incentive to produce polarising content (Rathje et al., 2021; Brady et al., 2023).

What would settle it is a design that manipulates the algorithm rather than observing it, holds the counterfactual constant across platforms and populations, and measures short-run attitudes alongside longer-run outcomes. Bakshy and colleagues separate exposure stages without randomising them, and Bail and colleagues randomise exposure but not the feed.

Rate the three positions by design. The small-effect position rests on two survey-and-traffic studies, one analysis of 10.1 million users that separates exposure stages and a 46-study meta-analysis (Guess et al., 2020b; Allcott and Gentzkow, 2017; Bakshy et al., 2015; Ferguson et al., 2025). The real-effect position rests on one four-week randomised deactivation, one natural experiment and one randomised exposure trial (Allcott et al., 2020; Braghieri et al., 2022; Bail et al., 2018). The incentive position rests on an analysis of 2,730,215 posts (Rathje et al., 2021). While the dispute is open, a designer should treat the ranking as a lever to be varied and measured.

Where the argument is heading

Brady and colleagues (2023) propose treating recommendation algorithms as part of the environment in which social learning happens, arguing they exploit biases towards prestigious, in-group, moral and emotional information, a pattern they call PRIME, which produces functional misalignment. Their remedies, bounded diversification and transparency of amplification, are design variables rather than education programmes.

Guess and colleagues (2020a) found a short media-literacy intervention raised discernment by 26.5 per cent in a representative US sample and 17.5 per cent in an educated Indian sample. The gain persisted in the United States but not in India, and the intervention did nothing in a rural northern Indian sample. An effect that holds in one population and vanishes in another is evidence about the conditions of the effect.

The sources this unit rests on

Each note states what the document did and found. Identifiers were resolved on 2 November 2026.

• Allcott, H., Braghieri, L., Eichmeyer, S. and Gentzkow, M. (2020). The welfare effects of social media. doi:10.1257/aer.20190658 Randomised deactivation of Facebook for four weeks before the 2018 US midterm election. It reduced political polarisation and factual news knowledge and raised subjective wellbeing, opposite to Braghieri, Levy and Makarin (2022).
• Allcott, H. and Gentzkow, M. (2017). Social media and fake news in the 2016 election. doi:10.1257/jep.31.2.211 Browsing data, fact-checking archives and a survey: fourteen per cent named social media their most important election news source, the average adult saw one or a few false stories, and just over half who recalled them believed them.
• Bail, C. A., Argyle, L. P., Brown, T. W., Bumpus, J. P., Chen, H., Hunzaker, M. B. F., Lee, J., Mann, M., Merhout, F. and Volfovsky, A. (2018). Exposure to opposing views on social media can increase political polarization. doi:10.1073/pnas.1804840115 Paid Twitter users to follow a bot posting opposing views for one month. Republicans who followed a liberal bot became substantially more conservative; the Democratic shift was slight and not significant.
• Bakshy, E., Messing, S. and Adamic, L. A. (2015). Exposure to ideologically diverse news and opinion on Facebook. doi:10.1126/science.aaa1160 Used deidentified data from 10.1 million US Facebook users to separate what friends shared, what the algorithm displayed and what users clicked. Individual choice limited cross-cutting exposure more than ranking did.
• Boulianne, S. (2015). Social media use and participation: a meta-analysis of current research. doi:10.1080/1369118x.2015.1008542 Meta-analysed 36 studies of social media use and participation. More than 80 per cent of coefficients were positive but only half significant; panel studies reported significant effects less often than cross-sectional surveys, and the campaign effect was minimal.
• Boxell, L., Gentzkow, M. and Shapiro, J. M. (2017). Greater Internet use is not associated with faster growth in political polarization among US demographic groups. doi:10.1073/pnas.1706588114 A polarisation index built from eight measures showed the fastest growth among the groups least likely to use the internet, and the model attributes a small share of the growth to it.
• Braghieri, L., Levy, R. and Makarin, A. (2022). Social media and mental health. doi:10.1257/aer.20211218 Exploited the staggered arrival of Facebook at US colleges. The rollout worsened student mental health and increased reported academic impairment, with mechanisms pointing to unfavourable social comparison.
• Brady, W. J., Jackson, J. C., Lindstrom, B. and Crockett, M. J. (2023). Algorithm-mediated social learning in online social networks. doi:10.1016/j.tics.2023.06.008 Reviewed how recommendation algorithms exploit social-learning biases towards prestigious, in-group, moral and emotional information, which the authors call PRIME, and propose bounded diversification and transparency of amplification.
• Guess, A., Lerner, M., Lyons, B., Montgomery, J. M., Nyhan, B., Reifler, J. and Sircar, N. (2020a). A digital media literacy intervention increases discernment between mainstream and false news. doi:10.1073/pnas.1920498117 Preregistered experiments in the United States and India. Discernment rose by 26.5 per cent in a representative US sample and 17.5 per cent in an educated Indian sample, persisted weeks later in the United States but not India, and failed in rural northern India.
• Guess, A. M., Nyhan, B. and Reifler, J. (2020b). Exposure to untrustworthy websites in the 2016 US election. doi:10.1038/s41562-020-0833-x Combined survey and web-traffic data from the 2016 campaign. Untrustworthy websites were a small share of information diets on average, largely consumed by a subset with strong pro-attitudinal preferences, so claims about prevalence have been overstated.
• Pennycook, G., McPhetres, J., Zhang, Y., Lu, J. G. and Rand, D. G. (2020). Fighting COVID-19 misinformation on social media. doi:10.1177/0956797620939054 Tested an accuracy prompt with more than 1,700 US adults. Asking people to judge an unrelated headline before sharing nearly tripled truth discernment in later sharing intentions.
• Pennycook, G. and Rand, D. G. (2019). Lazy, not biased: susceptibility to partisan fake news is better explained by lack of reasoning than by motivated reasoning. doi:10.1016/j.cognition.2018.06.011 Tested more than 3,400 participants on cognitive reflection and headline accuracy. Higher reflection predicted better discernment even for ideologically aligned headlines, so susceptibility tracked analytic engagement, not partisan motivation.
• Rathje, S., Van Bavel, J. J. and Van der Linden, S. (2021). Out-group animosity drives engagement on social media. doi:10.1073/pnas.2024292118 Analysed 2,730,215 posts from news outlets and political leaders. Out-group language was the strongest predictor of sharing: each out-group term raised the odds of a share by 67 per cent, and its effect was 4.8 times that of negative affect.
• Vosoughi, S., Roy, D. and Aral, S. (2018). The spread of true and false news online. doi:10.1126/science.aap9559 Analysed about 126,000 stories tweeted by about 3 million people more than 4.5 million times from 2006 to 2017, classified by six fact-checking organisations. Falsehood diffused farther, faster, deeper and more broadly than truth, and robots spread both at the same rate.
• Grinberg, N., Joseph, K., Friedland, L., Swire-Thompson, B. and Lazer, D. M. J. (2019). Fake news on Twitter during the 2016 U.S. presidential election. doi:10.1126/science.aau2706 Linked voter accounts to fake news sources and found engagement extremely concentrated, with 1 per cent of users accounting for 80 per cent of exposures. It separates a small average effect from a severe tail.
• Ferguson, C. J., Kaye, L. K., Branley-Bell, D. and Markey, P. (2025). There is no evidence that time spent on social media is correlated with adolescent mental health problems: findings from a meta-analysis. doi:10.1037/pro0000589 Pooled 46 studies of youth social media use and mental health, finding the pool unable to support claims of harm.