• Social media now shapes adolescent attention and emotion through engagement-driven algorithms, yet youth are expected to self-regulate in systems misaligned with their developmental capacities.
• Youth-safe default settings can redesign recommendation starting points, slowing harmful escalation and reducing reliance on individual self-control before habits and vulnerability patterns are established.
• Schools and families reinforce digital skills through repeated practice and modeling, helping adolescents translate awareness into action in environments where algorithmic systems alone can’t support regulation.

For many of today’s adolescents, social media has become a defining part of their daily lives, fueled by the rapid growth of platforms like Meta, Instagram, YouTube, Snapchat, Discord, and TikTok. As of 2024, more than five billion people worldwide hold a social media account, and that figure is expected to exceed six billion in the next few years.¹ Adolescents, unsurprisingly, account for a substantial share of this audience. Across a range of studies, most teenagers aged 13 to 17 report using at least one social media platform, with many reporting daily use of at least one platform.²

As reach continues to grow, time spent inside these systems has expanded as well. Recent surveys suggest adolescent girls aged 16 to 24 may spend more than three hours per day on social media, while boys of the same age group average closer to two and a half hours.² Time alone doesn’t determine harm, but it does signal how much attention is now routed through digital environments. What young people watch, compare, and return to may increasingly shape how information and social feedback are absorbed and remembered, especially when content is optimized to hold attention through intensity or novelty.

Concerns about mental health have grown alongside this rise. Meta-analytic evidence suggests a dose–response relationship between time spent on social media and depressive symptoms, with risk tending to increase as daily exposure rises.³ These findings don’t suggest that social media harms every adolescent, and they shouldn’t be read as deterministic. Still, they raise concerns about cumulative exposure during a developmental period marked by heightened emotional sensitivity and uneven regulatory capacity.⁴

Perhaps less apparent to young users, yet often shaping their experience in powerful ways, is algorithmic design. Platforms rely on artificial intelligence systems that track behavior continuously, recording where users pause, what they replay, and what draws them back to revise recommendations in real time. These systems aren’t passive. They’re trained on engagement signals and may amplify content through repetition and escalation. 

From an economic perspective, youth attention has become highly valuable. In the United States alone, children and adolescents have generated billions of dollars in advertising revenue in recent years,⁵ which may encourage design choices that extend use rather than support disengagement.

Taken together, adolescents aren’t only spending more time online, but they’re spending it inside systems that actively train attention through reward and repetition. When engagement drives revenue, exposure isn’t likely to slow on its own, even when stress, comparison, or distress begins to rise. That gap could leave developing users absorbing risks to emotional regulation and mental well-being; decision-science-informed design choices and governance approaches may help reduce those risks.

Social media platforms run on personalization systems that are optimized for engagement, meaning platforms monitor their users’ behaviors.⁶ They track how long a video’s been viewed, whether it’s replayed, how slowly someone scrolls, what gets saved, and how often users come back after closing the app. Those signals are collected across sessions and used to revise what appears next. In earnings reports and product reviews, time spent still shows up as the metric that counts.⁷

That logic sits squarely within attention economics.⁸ Attention is treated as a scarce resource, one that can be captured, measured, and reused. Recommendation systems rely on machine-learning methods, including reinforcement learning, to test what holds attention and then serve more of it. As new data comes in, models retrain and adjust to sustain user engagement. 

This is where concerns about youth well-being have begun to rise. Reports from Amnesty International and Reset Australia suggest these systems could expose young people to untrusted, unverified, and radical material, with consequences for peer relationships and mental health.^9,10 Adolescents may be especially vulnerable to echo chambers and the spread of misogynistic content, disordered eating content, or sensational or violent media.^11,12,13 These exposures don’t usually require deliberate searching; content that racks up views, longer watch time, or replays is treated as successful, so similar material keeps resurfacing. Even brief engagement can train feeds, and over time, repetition may normalize what began as incidental exposure.

Evidence from England and Wales has shown how this process can develop. A mixed-methods study combined long-form interviews with young people, algorithmic analysis of over 1,000 social media videos, and interviews and roundtable discussions with school leaders.¹³ Together, these data sources were used to examine how digital environments encourage and normalize harmful ideologies and how such exposure affects young people’s well-being. The study reported three key findings: recommendation systems increased exposure to radical material, misogynistic content was often presented as entertainment, and hateful tropes emerged in young people’s language and behavior, with negative downstream consequences for peer relationships and mental health.

A separate line of evidence comes from research on eating disorders and social media use. In a month-long observational study, researchers analyzed 1.03 million TikTok videos delivered to 42 individuals with eating disorders and 49 healthy controls.¹⁴ Four categories of content relevant to eating-disorder psychopathology were examined: appearance-oriented videos, dieting videos, exercise videos, and toxic eating-disorder content. Algorithms delivered substantially more of all four categories to users with eating disorders, including a 4,343% increase in toxic eating-disorder content. These delivery patterns were only weakly correlated with users’ active “likes.” Instead, algorithmic delivery itself was more strongly associated with symptom severity, suggesting that personalization processes may exacerbate vulnerability independent of deliberate engagement.

The evidence therefore suggests that recommender systems may amplify content that aligns with vulnerability, including misogynistic narratives and eating-disorder material. Particularly when harmful content is framed or perceived as entertainment, higher engagement makes it more likely to recur in recommendation feeds.

If recommender systems determine what content gains momentum, youth safety sets the boundaries within which those systems operate within. Reducing risk depends on how platforms structure defaults, reward engagement, and evaluate the consequences of those choices over time.

Default Safety: Redesign the Starting Point, Not the User

One practical way to reduce harm at scale may be to adjust starting conditions rather than rely on downstream correction. Defaults, or pre-set options that take effect when no active choice is made, can be a simple but powerful design tool because they shape behavior before reflection or effort enters the picture. Platforms already use defaults to guide what users see, click, and return to, and youth safety could operate through the same mechanism.

If youth accounts opened in safety-optimized modes, early exposure might look meaningfully different. Extreme personalization could be delayed until enough neutral behavioral data has been collected, and topic diversity could be increased so a single theme doesn’t dominate a feed.

In practice, limiting repeat exposure would mean placing caps on how often similar content is recommended within a short period. If a youth user watches several videos from the same theme in a row, the system could temporarily stop recommending more of that theme and switch to unrelated topics. Content from accounts the user already follows would still appear, but the algorithm wouldn’t continue pushing similar material through “For You” or “Explore” feeds. The goal isn’t to block content, but to slow repetition before patterns harden.

Align Platform Incentives With Youth Safety

Behavior tends to change when incentives change. Decision science tells us that actors often adjust when costs and benefits make safer options the rational path. At present, engagement metrics carry that weight because they map directly onto revenue and valuation. Youth safety, by contrast, often sits outside of core performance indicators.

That imbalance could be addressed through aligned incentives. Procurement and advertising conditions offer one route. Governments, educational bodies, NGOs, and large advertisers might require participation in audited youth-safety programs or standardized algorithmic risk reporting as a condition of partnership. Platforms that demonstrate safer recommender behavior could gain access to contracts or campaigns to which others may not.

Regulatory credits provide another lever. Similar to mechanisms used in environmental policy, platforms that meet defined youth-safety performance thresholds might receive compliance credits, reduced fees, or faster approvals for related services. These approaches introduce upside rather than relying only on penalties. Over time, that upside could influence strategic priorities. When safety begins to affect access to markets, partnerships, or approvals, it’s more likely to enter routine governance. 

Independent Algorithmic Audits: Make Harm Legible

Recommender systems operate inside fast feedback loops. Engineers run experiments, models learn from behavior, and performance is judged using metrics such as watch time and retention. That loop allows systems to adapt quickly. It may also amplify vulnerability quickly.

Independent audits could help by adding an external reference point, introducing measures that sit outside engagement metrics and internal experimentation. Under the EU’s Digital Services Act, very large platforms already conduct annual systemic risk assessments that include algorithmic impacts on minors and well-being.^15,16 That requirement could be expanded or standardized across nations. Audits could examine exposure concentration, repeat amplification, and escalation speed following initial engagement. When those dynamics are tracked and reported, they’re harder to dismiss as edge cases.

The goal isn’t disclosure for its own sake, it’s operational visibility. When youth risk appears alongside privacy risk, legal exposure, or brand risk, leadership is more likely to manage it deliberately. Product teams may surface concerns earlier. Executives might weigh trade-offs sooner.

Everyday social media use depends on a surprisingly complex set of skills. Users are expected to regulate emotional reactions, interpret social feedback, infer why certain content appears, decide when to disengage, and resist design features that reward continued attention.¹⁷ These demands aren’t trivial for adults, and they may be even more taxing for adolescents, whose regulatory and interpretive systems are still developing. 

Young users are often navigating high-stimulation environments where control mechanisms aren’t fully stable and buffers against cognitive overload don’t always hold.¹⁸ Developmental science has long described adolescence as a prolonged and uneven process rather than a brief, linear transition toward adulthood. Neurological, emotional, and social systems continue changing into the mid-twenties; during this period, emotional responsiveness often increases faster than executive control.¹⁸ Adolescents may react more strongly to social cues, recover more slowly from rejection, and struggle more with impulse regulation than adults do.¹⁹ These patterns don’t emerge at a single age, and they aren’t distributed evenly across individuals. 

Digital environments haven’t been built around that timeline. Adolescents are expected to manage algorithmically curated feeds, public metrics of approval, and constant opportunities for comparison. They must decide what to watch, how to interpret feedback, and when to log off, even when signals are persistent, quantified, and socially visible. When distress shows up, responsibility is often framed as a matter of individual use, even though the surrounding environment may be reinforcing continued engagement.

Emotion regulation makes this mismatch visible, but it’s only one part of the picture. Regulating emotion involves pausing a reaction, reinterpreting what’s happening, and choosing when to disengage.¹⁹ Those abilities are still stabilizing during adolescence, yet they’re needed constantly online. At the same time, adolescents are also expected to infer algorithmic intent, recognize patterns of repeated exposure, and distinguish between organic popularity and amplified content. Those interpretive skills tend to develop through experience, not instruction, and they may be uneven across youth.

Experimental studies on social media feedback illustrate how these demands can converge. In one study, adolescents exposed to lower engagement during standardized online interactions reported stronger feelings of rejection and more negative self-evaluations than peers who received higher feedback.²⁰ Greater sensitivity to low engagement was also associated with higher depressive symptoms at follow-up. These findings don’t suggest a single pathway to harm, but they highlight how emotionally salient feedback, repeated exposure, and limited disengagement skills can interact over time.

The bigger challenge, then, isn’t simply that adolescents feel more when they’re online. It’s that everyday social media use assumes adult-level emotional regulation, self-control, and interpretive capacity in users who are still developing those systems. Some adolescents may navigate these demands without major difficulty. Others may face disproportionate risk when multiple skill gaps compound at once.

Digital platforms place increasing demands on emotional regulation, judgment, and self-management. These skills are still developing in adolescence, and schools and families remain the settings where they’re most often learned and practiced. Unlike platforms, these contexts already shape how young people interpret social feedback, manage stress, and revise behavior after mistakes.^21,22 The opportunity lies in strengthening that scaffolding so it better supports digital use, rather than assuming young users will adapt on their own.

Schools may offer one of the most consistent points of access. School-based programs reach most adolescents by default and can engage students who wouldn’t actively seek help. They also allow educators to work with groups that may already face elevated risk without requiring individual identification. Evidence across health and mental-health research suggests that school-based interventions can reduce symptoms of anxiety and depression, improve academic performance, and enhance overall well-being.^23,24 These effects tend to emerge when students repeatedly practice skills in shared settings, rather than through single lessons delivered in isolation.²⁴

This logic has extended into social media–specific programs, though results remain mixed. A recent cluster randomized controlled trial evaluated SoMe, a classroom-based social media literacy program focused on body image, dieting, and wellbeing. Nearly 900 students aged 11–15 completed weekly sessions over four weeks.²⁵ At six-month follow-up, improvements appeared in dietary restraint and depressive symptoms among girls, while effects for boys were smaller. The findings were preliminary, but they suggest an important constraint: universal programs may help some students more than others, and literacy alone may not generalize evenly across developmental or social contexts.

Families represent a second, complementary institution. Many parents report concern about online content, yet survey data show that awareness drops sharply outside traditional media. While roughly 80% of parents report high awareness of what their child watches on television, fewer than 60% report the same for online video, and closer to 40% for social media.²⁶ That gap may limit opportunities for timely guidance.

Active co-use and co-viewing could help close that gap. When parents actively watch, scroll, or play alongside their children, they gain first-hand visibility into what appears, how fast it accumulates, and which themes repeat. Patterns that might otherwise stay hidden start becoming observable: a feed may be narrowing, comparison may be escalating, emotional reactions may be lasting longer than expected. Those moments create practical openings to pause, label what’s happening, and decide together whether to continue or step away.

Parental mediation theory explains why this process can be effective.²⁷ The theory proposes that youth outcomes are shaped through ongoing communication around media use, rather than solely relying on restrictions. Skills develop when adults explain how content is selected, discuss emotional responses as they occur, and revisit decisions after exposure. Over time, adolescents may begin connecting their own actions with changes in what appears next. 

Social learning theory strengthens this account. Children and adolescents are constantly observing adult behavior, especially in ambiguous situations.²⁸ When parents turn autoplay off, step away after emotional spikes, or limit repeat exposure, they’re demonstrating regulation in real conditions. That modeling could reinforce habits more effectively than verbal advice alone, because it shows what disengagement actually looks like. Those behaviors may be shaping expectations about when attention should stop rather than continue.

One important caveat is that safety defaults can produce unintended spillovers if they’re applied too broadly. When platforms suppress categories like “dieting” or “thinspiration” to reduce harm, legitimate content may be caught in the net. Nutrition education, diabetes management, or neutral fitness communities could be demoted alongside genuinely harmful material. Over time, young users may infer that information is being hidden, which can increase curiosity rather than reduce it.

A second caveat concerns uneven implementation capacity. School-based digital well-being programs might look effective on paper but play out very differently across a range of settings. In well-resourced schools, for example, counselors could co-teach lessons, class sizes may be smaller, and students can practice the skills they’ve learned repeatedly over time. In under-resourced schools, the same curriculum may have been delivered as a one-off assembly or a rushed worksheet, limiting its influence. When that happens, effect sizes shrink, and the students who could benefit most receive the weakest version. This creates an equity risk: interventions intended to protect youth can widen gaps if differences in training, time, and support aren’t addressed up front.

Designing safer social media environments with youth development in mind can directly support emotion regulation, reduce repeated exposure to harmful content, and make disengagement easier when stress or social comparison escalates. That work doesn’t hinge on a single fix. It depends on how platforms set defaults, how risks are measured and reported, and how schools and families are equipped to reinforce digital skills in everyday settings.

Transparent algorithmic risk reporting could help make exposure patterns visible rather than inferred. Safer default settings could slow escalation before habits set in. At the same time, schools and parents remain essential partners, providing repeated practice, guidance, and modeling that platforms can’t offer on their own. When those roles are treated as complementary rather than separate, youth may be better supported as they learn when to engage, when to pause, and when to step back.

At The Decision Lab, we work with governments and organizations to apply behavioral science principles and theories where systems shape behavior at scale. If you’re exploring how to redesign digital environments, strengthen governance, or support youth development more effectively, we’d welcome the conversation. Let’s work together to build social media systems that help young people regulate, reflect, and recover—by design.

Digital Nudging

Digital nudging is the use of interface design in online spaces to steer behavior. Rooted in behavioral science, it includes cues like infinite scroll, push notifications, default settings, reminder emails, and even the exact wording of prompts. We usually notice nudges when they keep us scrolling, but the same tools can be used for good too. 

How to Use Social Media as a Vehicle for Accurate Health Information

Social media often gets framed as the enemy of mental health. There’s evidence linking heavier use to higher anxiety, depression, stress, self-injury, and suicidality in young people. Still, that’s not the whole story. These platforms can also be designed to surface accurate health information when people need it. Click here to learn what this type of design can look like in practice.

1. Statista. (2024). Number of social media users worldwide from 2017 to 2028 (in billions). Statista. https://www.statista.com/statistics/278414/number-of-worldwide-social-network-users/
2. Ahmed, O., Walsh, E. I., Dawel, A., Alateeq, K., Espinoza Oyarce, D. A., & Cherbuin, N. (2024). Social media use, mental health and sleep: A systematic review with meta-analyses. Journal of Affective Disorders, 367, 701–712. https://doi.org/10.1016/j.jad.2024.08.193 
3. Liu, M., Kamper-DeMarco, K. E., Zhang, J., Xiao, J., Dong, D., & Xue, P. (2022). Time spent on social media and risk of depression in adolescents: A dose–response meta-analysis. International Journal of Environmental Research and Public Health, 19, 5164. https://doi.org/10.3390/ijerph19095164 
4. Guyer, A. E., Silk, J. S., & Nelson, E. E. (2016). The neurobiology of the emotional adolescent: From the inside out. Neuroscience & Biobehavioral Reviews, 70, 74–85. https://doi.org/10.1016/j.neubiorev.2016.07.037 
5. Raffoul, A., Ward, Z. J., Santoso, M., Kavanaugh, J. R., & Austin, S. B. (2023). Social media platforms generate billions of dollars in revenue from U.S. youth: Findings from a simulated revenue model. PLOS ONE, 18(12), e0295337. https://doi.org/10.1371/journal.pone.0295337 
6. Mann, S., Pathak, N., Alhussein, M., Prajapati, J., Aurangzeb, K., & Sharma, N. (2025). Enhancing social media user engagement through personalized content classification. Contemporary Mathematics, 1076–1090. https://doi.org/10.37256/cm.6120255700 
7. Snap Inc. (2025, August 5). Snap Inc. announces second quarter 2025 financial results [Press release]. https://investor.snap.com/news/news-details/2025/Snap-Inc--Announces-Second-Quarter-2025-Financial-Results/default.aspx
8. Loewenstein, G., & Wojtowicz, Z. (2025). The economics of attention. Journal of Economic Literature, 63(3), 1038–1089. https://doi.org/10.1257/jel.20241665 
9. Amnesty International. (2024). Driven into darkness. https://www.amnesty.org/en/documents/pol40/7350/2023/en/
10. Reset Australia. (2022). Algorithms as a weapon against women: How YouTube lures boys and young men into the ‘manosphere’. https://www.isdglobal.org/wp-content/uploads/2022/04/Algorithms-as-a-weapon-against-women-ISD-RESET.pdf 
11. Harriger, J. A., Evans, J. A., Thompson, J. K., & Tylka, T. L. (2022). The dangers of the rabbit hole: Reflections on social media as a portal into a distorted world of edited bodies and eating disorder risk and the role of algorithms. Body Image, 41, 292–297. https://doi.org/10.1016/j.bodyim.2022.03.008 
12. Youth Endowment Fund. (2024, November 25). 70% of teens see real-life violence on social media, reveals new research. https://youthendowmentfund.org.uk/news/70-of-teens-see-real-life-violence-on-social-media-reveals-new-research/
13. Regehr, K., Shaughnessy, C., Zhao, M., Cambazoglu, I., Turner, A., & Shaughnessy, N. (2025). Normalizing toxicity: The role of recommender algorithms for young people’s mental health and social wellbeing. Frontiers in Psychology, 16, 1523649. https://doi.org/10.3389/fpsyg.2025.1523649 
14. Griffiths, S., Harris, E. A., Whitehead, G., Angelopoulos, F., Stone, B., Grey, W., & Dennis, S. (2024). Does TikTok contribute to eating disorders? A comparison of the TikTok algorithms belonging to individuals with eating disorders versus healthy controls. Body Image, 51, 101807. https://doi.org/10.1016/j.bodyim.2024.101807 
15. European Commission, Directorate-General for Communications Networks, Content and Technology. (2025, December 12). DSA: Very large online platforms and search engines. Shaping Europe’s Digital Future. https://digital-strategy.ec.europa.eu/en/policies/dsa-vlops
16. EUR-Lex. (n.d.). Digital Services Act: Auditing very large online platforms and search engines (summary). https://eur-lex.europa.eu/EN/legal-content/summary/digital-services-act-auditing-very-large-online-platforms-and-search-engines.html
17. Liu, T. Y., Ko, W. T., Griffiths, M. D., Pakpour, A. H., Üztemur, S., Ahorsu, D. K., & Lin, C. Y. (2024). The associations between levels of inattention/hyperactivity and social media addiction among young adults: The mediating role of emotional regulation strategies in self-blame and blaming others. Acta Psychologica, 251, 104338. https://doi.org/10.1016/j.actpsy.2024.104338 
18. Dumontheil, I. (2016). Adolescent brain development. Current Opinion in Behavioral Sciences, 10, 39–44. https://doi.org/10.1016/j.cobeha.2016.04.012 
19. Silvers, J. A., McRae, K., Gabrieli, J. D., Gross, J. J., Remy, K. A., & Ochsner, K. N. (2012). Age-related differences in emotional reactivity, regulation, and rejection sensitivity in adolescence. Emotion, 12(6), 1235–1246. https://doi.org/10.1037/a0028297 
20. Lee, H. Y., Jamieson, J. P., Reis, H. T., Beevers, C. G., Josephs, R. A., Mullarkey, M. C., & Yeager, D. S. (2020). Getting fewer “likes” than others on social media elicits emotional distress among victimized adolescents. Child Development, 91(6), 2141–2159. https://doi.org/10.1111/cdev.13451 
21. Farley, J. P., & Kim-Spoon, J. (2014). The development of adolescent self-regulation: Reviewing the role of parent, peer, friend, and romantic relationships. Journal of Adolescence, 37(4), 433–440. https://doi.org/10.1016/j.adolescence.2014.03.016 
22. Pedrini, L., Meloni, S., Lanfredi, M., & Rossi, R. (2022). School-based interventions to improve emotional regulation skills in adolescent students: A systematic review. Journal of Adolescence, 94(8), 1051–1067. https://doi.org/10.1002/jad.12090 
23. Grande, A. J., Hoffmann, M. S., Evans-Lacko, S., Ziebold, C., de Miranda, C. T., McDaid, D., & Ribeiro, W. S. (2023). Efficacy of school-based interventions for mental health problems in children and adolescents in low- and middle-income countries: A systematic review and meta-analysis. Frontiers in Psychiatry, 13, 1012257. https://doi.org/10.3389/fpsyt.2022.1012257 
24. Durlak, J. A., Weissberg, R. P., Dymnicki, A. B., Taylor, R. D., & Schellinger, K. B. (2011). The impact of enhancing students’ social and emotional learning: A meta-analysis of school-based universal interventions. Child Development, 82(1), 405–432. https://doi.org/10.1111/j.1467-8624.2010.01564.x 
25. Gordon, C. S., Jarman, H. K., Rodgers, R. F., McLean, S. A., Slater, A., Fuller-Tyszkiewicz, M., & Paxton, S. J. (2021). Outcomes of a cluster randomized controlled trial of the SoMe social media literacy program for improving body image–related outcomes in adolescent boys and girls. Nutrients, 13(11), 3825. https://doi.org/10.3390/nu13113825 
26. Gabrielli, J., Marsch, L., & Tanski, S. (2018). TECH parenting to promote effective media management. Pediatrics, 142(1), e20173718. https://doi.org/10.1542/peds.2017-3718 
27. Jiow, H. J., Lim, S. S., & Lin, J. (2017). Level up! Refreshing parental mediation theory for our digital media landscape. Communication Theory, 27(3), 309–328. https://doi.org/10.1111/comt.12109 
28. Muuss, R. E. (1976). The implications of social learning theory for an understanding of adolescent development. Adolescence, 11(41), 61–78.