Less harmful social mediaCopy wiki link

The strongest alternative to excluding minors from social media is not one parental-control setting or one literacy lesson. It is to redesign the service around relationships, expression, and deliberate choice rather than continuous engagement.

This approach starts from a more precise empirical claim. Social media is not one exposure, and neither “time online” nor the user’s age identifies the harm mechanism. Sleep-displacing notifications, appearance comparison, stranger contact, harassment, commercial profiling, and recommendation spirals are different problems. They require different interventions.

Social media has costs and functionsCopy wiki link

The JRC social media and youth mental health review found no uniform relationship between time or frequency of use and poor mental health. Effects were mixed, usually small, and dependent on the person, activity, content, and outcome measured. More consistent concerns involved problematic use, social comparison and appearance pressure, sleep disruption, and harmful encounters.

That distinction matters for policy. A generic reduction in use may leave the causal feature intact, move activity to another channel, or remove beneficial use alongside harmful use. One randomized restriction study found that university students substituted messaging for restricted social networks without significant gains in well-being or academic outcomes.¹ It is not a study of younger adolescents, but it demonstrates why displacement must be measured rather than assumed away.

Young people use social media for friendship, social support, identity exploration, sexual and political expression, creative production, information, community, and participation. These benefits are not reasons to ignore harms. They are outcomes an intervention should try to preserve and costs that a blanket age gate must count.

Match each mechanism to a design responseCopy wiki link

The EU DSA protection of minors guidelines distinguish content, conduct, contact, consumer, and cross-cutting risks. The same separation produces a more useful design agenda:

Each intervention remains a testable proposal whose outcome measure follows from the mechanism it claims to change. “Keep under-16s out” is harder to interpret: any observed effect mixes lost harms, lost benefits, circumvention, and movement to other services.

Replace engagement-first architectureCopy wiki link

Many platforms have a structural conflict. Their revenue increases when people return frequently, consume more recommendations, and generate more behavioral data. The same system can make protective settings cosmetic while the ranking and notification machinery still pushes engagement.

The European Commission’s 2026 preliminary findings against TikTok, Instagram, and Facebook focus on this architecture: infinite scroll, autoplay, push notifications, and highly personalized recommendations.^{2 3} These are preliminary enforcement views, not final legal findings, but they identify features that can be changed without excluding young people from communication.

Children describe the machinery’s effect directly. In United Kingdom surveys, 59% reported staying up late on devices and 46% continuing to watch or play when no longer enjoying it, with infinite scroll and streaks named as the hooks, and both children and parents called time online their most immediate day-to-day concern.⁴

A relationship-first service would instead make the user choose people, communities, or topics deliberately. It would provide bounded sessions and visible stopping points, separate recent posts from algorithmic suggestions, and let the user inspect, reset, or disable the inference profile. Success metrics would include unwanted exposure, sleep disruption, harassment, regretted use, and durable social benefit, not merely time spent or return frequency.

This could require a different business model. Subscription, public-service, cooperative, and contextual-advertising models change what the business is paid for, which lowers the reward for maximizing surveillance and attention without guaranteeing anything about the design itself. Interoperability and data portability can also lower the cost of leaving a harmful service without losing every relationship.

Federated, non-commercial networks remove the incentive itselfCopy wiki link

Changing what a single company is paid for is one lever. A structurally different proposal removes the single company: a federated network of independently operated, often volunteer-run servers using a shared protocol such as ActivityPub, with no central entity that profits from maximizing attention and therefore no organizational incentive to build an engagement-optimizing recommender at all. This is a governance and architecture change, not merely a different revenue source layered onto the same design, and it is worth separating explicitly from the business-model paragraph above: a subscription-funded but still centrally operated platform can still choose to maximize session length to justify its price, while a server with no attention-based revenue model has nothing to optimize engagement toward.

The trade is real rather than free. Federated networks typically distribute moderation and safety-default decisions to each server’s operators instead of one company’s trust-and-safety organization, which can produce inconsistent protections for minors across instances, uneven resourcing for urgent response, and a discovery experience that is less centrally legible to regulators applying the DSA’s platform-level obligations. ActivityPub’s audience fields are access-control instructions to servers rather than end-to-end encryption, and remote copies on federated instances can outlive an author’s deletion request. A federated alternative to a commercial platform therefore does not inherit the same accountability mechanisms by default; it needs its own answer for confidential reporting, urgent escalation, and enforcement consistency across independently operated servers, not merely the absence of a profit motive.

This proposal has not received the same empirical evaluation in this note as contact controls or literacy interventions. No study was found directly comparing harm outcomes for minors on a federated, non-algorithmic network against a commercial engagement-optimized one. The case for it here is structural and mechanistic rather than outcome-tested: removing the entity whose revenue depends on maximizing attention removes a cause the JRC social media and youth mental health review and the Commission’s TikTok and Meta preliminary findings both identify as central to compulsive use and harmful recommendation spirals,²³ without requiring that any single company be persuaded to act against that incentive voluntarily. Testing this proposal against the same outcomes this note applies to contact and architecture controls is part of the Which social-media designs reduce harm for young users research agenda.

Safer defaults are useful but incompleteCopy wiki link

Ofcom protective social media defaults trial found that 94% of adults and participants aged 13 to 17 retained protective network, location, and messaging defaults in a simulated platform. This shows that protective defaults can be usable and persistent. The study did not expose people to a functioning social network or measure whether harms fell.

Defaults should therefore be treated as an implementation tool, not the outcome. Platforms should test whether they reduce unwanted contact, harmful repetition, sleep loss, or financial loss in real use. Independent researchers need access to the necessary platform data, and evaluations should publish effects by age, gender, vulnerability, and use pattern rather than only one average.

The same caution applies to literacy. Teaching users how ranking, editing, advertising, and social comparison work can support autonomy. One cluster-randomized social-media literacy intervention, however, did not improve the main body-image and well-being outcomes across its full sample; some benefits appeared only for girls at later follow-up.⁵ Literacy should complement changes to the environment rather than excuse a deliberately manipulative environment.

Recognition education and trusted relationships as a distinct alternativeCopy wiki link

Contact restriction is not the only alternative to a platform-wide age gate for harassment and grooming. A second family of interventions works upstream of any single platform: it builds the adolescent’s own capacity to recognize manipulation, and the relationships through which an unwanted encounter gets disclosed and interrupted early. This family deserves the same evidentiary treatment this note gives to contact and architecture controls, rather than the token mention literacy work often receives as a caveat to product design.

Brief, targeted grooming-recognition education has direct trial evidence. A Spanish RCT of 856 adolescents aged 11 to 17 randomized students to a sub-hour online educational intervention about grooming tactics or to an unrelated resilience-building control. Adolescents who received the grooming intervention were significantly less likely to respond to a later sexual solicitation from an adult with sexualized interaction such as sending photos, and their knowledge of grooming tactics rose and was retained at three- and six-month follow-up.⁶ This is a randomized behavioral outcome, not merely a knowledge-quiz result, and it directly rebuts the idea that education can only raise awareness without changing what happens during an actual solicitation.

The broader school-based child sexual abuse prevention literature is older and larger but more equivocal. A Cochrane review of 24 randomized and quasi-randomized trials (5,802 children) found consistent gains in protective knowledge and self-protective behaviors that were retained at six months, with no measurable increase in child anxiety or fear.⁷ Its disclosure finding is genuinely mixed: an unadjusted meta-analysis favored higher disclosure rates among children who received the intervention, but the effect lost statistical significance once the analysis corrected for clustering within classrooms and schools, a correction more than half the included trials had originally omitted. The honest reading is that these programs reliably teach concepts and behavioral scripts, and plausibly help some children disclose sooner, but the evidence for a disclosure effect is weaker than the evidence for a knowledge or behavior effect, and no trial in the review could measure whether the programs actually reduced completed abuse.

Most disclosure goes to peers, not adults or platformsCopy wiki link

The contact-restriction model in the table above implicitly assumes that a harassed or groomed adolescent’s live options are enduring the contact, using a platform reporting tool, or reaching a parent or authority. Population survey evidence says adolescents mostly do something else first: they tell a friend.

A pan-European study of children’s coping after online risk exposure found that most adolescents in high-internet-access countries respond to a risky encounter with a positive strategy such as seeking help from friends, or a neutral strategy such as ignoring it, and that “most strategies tend to exclude adult involvement.”⁸ A Finnish population-based study of disclosed sexual violence among 11- to 17-year-olds found that victims told a peer in 63% of cases, a parent in 23%, and an authority in only 10%, regardless of whether the offender was a stranger or known to the victim; shame, distrust of adults, and disbelief that disclosure would help were the reasons victims gave for staying silent.⁹ A US national survey of adolescents who had experienced peer harassment found that bystanders who witnessed it and intervened usually did so through direct peer support or defense, and that adolescents rated bystander support as more helpful than most adult or institutional responses.¹⁰

This evidence reframes what “response to harassment” should optimize for. A design agenda built only around confidential reporting to the platform and escalation to authorities is optimizing for the disclosure path adolescents use least. The stronger design lever is the peer relationship itself: whether the adolescent’s ordinary friends know how to recognize a harmful pattern, respond supportively, and help escalate when needed, and whether at least one adult exists who the adolescent trusts enough to approach voluntarily rather than being forced into disclosure through platform surveillance.

Peer bystander training has direct experimental evidence in this domain. A Belgian cluster-randomized trial of Friendly Attac, a serious digital game teaching prosocial bystander responses to cyberbullying, increased self-efficacy and positive bystander intentions among young adolescents.¹¹ Results are not uniform: a Norwegian cluster-randomized trial of a self-help app for adolescents experiencing cyberbullying, NettOpp, found no significant intention-to-treat effect on mental health outcomes, illustrating that a digital tool’s availability does not guarantee real-world uptake or effect, distinct from the question of whether recognition and disclosure skills help once they are actually used.¹² A Swedish study of bystander behavior in peer victimization found that adolescents with higher defender self-efficacy and better student-teacher relationship quality were more likely to act as defenders rather than reinforcers or bystanders, identifying relationship quality with a trusted adult as a measurable predictor of protective peer behavior, not only a soft precondition for it.¹³

Contact hardening has real but narrower displacement evidence than assumedCopy wiki link

Message limits, adult-minor interaction constraints, and screenshot prevention could plausibly push offending behavior toward channels that are harder to monitor and intervene in, rather than eliminating it. That is a hypothesis about Displacement, and it has real evidentiary support in one specific domain and none yet demonstrated in another. The base rate is worth stating before the specifics: across situational crime prevention, displacement appears in about a quarter of the observations that test for it, the opposite spillover appears about as often, and where relocation does occur it is usually smaller than the intervention’s own effect.

The evidence is strongest for content-detection displacement. A survey of more than 30,000 self-reported offenders by the Finnish organization Protect Children found that many deliberately chose end-to-end encrypted (E2EE) platforms because they perceived a lower risk of detection or prosecution.¹⁴ After Meta enabled E2EE by default on Messenger in December 2023, reports of child sexual abuse material (CSAM) to the US NCMEC CyberTipline fell by around 40% in a single year, a drop attributed to reduced platform-side detection capability rather than falling abuse.¹⁵ This is genuine evidence that hardening one control point, automated content scanning, shifts offending toward channels where that control point does not reach.

No comparably direct evidence was found that hardening contact controls specifically, such as message limits or adult-minor interaction restrictions, produces the same displacement for grooming initiation. The mechanism is plausible by analogy, and Whittle, Hamilton-Giachritsis, and Beech’s interviews with groomed adolescents describe offenders adapting their approach to each platform’s specific features,¹⁶ but this is offenders adapting tactics within a platform, not documented migration to a less accountable platform in response to contact-limiting features specifically. Treat contact-control displacement as a plausible, underexamined risk rather than a demonstrated one, and design its evaluation into any contact-restriction rollout rather than assuming either that it will or will not occur.

A blanket adult-minor contact restriction selects against its own purposeCopy wiki link

A separate and more specific problem affects adult-minor interaction constraints in particular, distinct from the offender-displacement question above. The control only does what it claims if the platform can reliably sort adults from minors. An independent technical assessment of deployed age-assurance architectures found that every current age signal, including self-declaration, government-ID checks, and facial age estimation, suffers from either accuracy or availability limits, and concluded that “it is not technically feasible to build an age assurance system which would prevent all minors from accessing restricted content or experiences without also blocking large numbers of adult users.”¹⁷ The same report found that server-based enforcement is circumventable with a VPN, device-based enforcement is circumventable with a non-enforcing device, and several mechanisms can be defeated by cooperation with a consenting adult, such as an adult unlocking a device or lending payment credentials to a minor who wants access. These are general weaknesses in the age signal itself, usable by whichever party has a reason to defeat it; the report does not isolate adult-side circumvention specifically, so treat “adults can defeat these same mechanisms” as the same class of weakness applied to the other direction, not as a separately measured finding.

This matters for adult-minor interaction constraints specifically because a leaky age boundary does not fail symmetrically. A constraint that limits which adults can message which minors creates two adult populations with different incentives to defeat it. An adult with no sexual interest in minors has little reason to spend effort bypassing an age check whose only effect is blocking an interaction they were not seeking anyway. An adult seeking to groom a minor has a strong, specific motive to defeat the same check, and the verification failure modes above (VPNs, unenforced devices, borrowed credentials, misreported age, a false or absent parental-consent step) are exactly the tools a motivated adult can use deliberately where an uninterested adult would not bother. The plausible net effect is not that the constraint stops grooming and leaves ordinary adult-minor interaction unaffected. It is that the constraint disproportionately removes non-motivated adults from a minor’s ordinary online contact pool while a motivated offender routes around it, which is the reverse of routine activity theory’s guardianship logic: crime is more likely where a motivated offender meets a suitable target in the absence of a capable guardian,¹⁸ and a constraint that filters out capable but non-motivated adults more effectively than it filters out motivated ones removes guardianship capacity from the interaction rather than adding it.

This causal chain has not been directly tested for social-media adult-minor contact rules specifically, and it should be treated as a documented mechanism assembled from adjacent evidence rather than a demonstrated result. Each of its links has independent support: verification is genuinely leaky,¹⁷ motivated and non-motivated actors face asymmetric incentives to defeat it, and guardianship presence is a recognized protective factor in the general routine-activity literature.¹⁸ What is missing is a study that follows adult-minor contact restrictions specifically into measured changes in ambient adult presence, grooming initiation, or intervention rates, as opposed to studies of general online-safety guardianship, where one existing study found that guardianship-improving measures were the least effective of the routine-activity factors tested against general online victimization, a genuinely discouraging result for guardianship-based reasoning that this note’s argument should not paper over.¹⁹ This is a case where the honest position is that the mechanism is coherent and worth researching directly, not that it is already established.

Minor-only enclaves can concentrate risk rather than remove itCopy wiki link

A related but distinct proposal, common in practice on Discord and similar platforms, excludes adults from a space entirely rather than restricting specific adult-to-minor messages within a mixed-age platform. A server, community, or app markets itself as being for minors only and removes or bans adult members, sometimes through age-gating at the door and sometimes through informal community norms. The composition question is different from the contact-filter question above: it is not whether a motivated adult can reach a specific minor, but who ends up ambiently present in the space at all.

The same asymmetric-incentive logic applies with the roles reversed. A parent, teacher, older sibling, youth worker, or other adult with a legitimate reason to be present has comparatively little motivation to falsify their age or evade a minor-only space’s entry check, since the space is not built to serve their interests and evading the check gains them nothing they were seeking. A motivated groomer has every reason to falsify downward or otherwise defeat the same check to gain entry, including posing as a minor. Discord grooming cases document offenders using voice-altering software specifically to pose as a peer to their targets,²² and predators identifying vulnerable minors on one youth-heavy platform before migrating them to a second, less supervised one.²³ A minor-only space can therefore end up with fewer capable adult guardians than a mixed-age space, not more, while offering no comparable reduction in the motivated offenders who successfully enter it.

This is not a hypothetical failure mode. The extortion network known as “764” operated across Discord, Minecraft, and Roblox servers built for children, and academic analysis of victim accounts describes the group as deliberately infiltrating platforms and spaces “designed for their use and enjoyment,” extorting minors into self-harm and producing child sexual abuse material with little adult oversight intervening before serious harm occurred.²⁴ This is one documented case rather than a general estimate of how often minor-only spaces host this dynamic, and it does not establish a base rate; it establishes that the mechanism is not merely theoretical.

The general protective-factor evidence supports the underlying logic without testing the enclave case directly. A time-diary study of adolescent substance use found that the absence of authority figures functioned as an independent risk factor for deviant behavior, alongside unstructured activity and peer presence, regardless of whether the other risk factors were also present.²⁵ A qualitative study of Dutch minors with experience of commercial sexual exploitation found that social isolation, withdrawal from stable relationships, and distrust of formal institutions and support figures were recurring features of the contexts in which exploitation occurred, while supportive relationships across the young person’s social ecology were protective.²⁶ Neither study measured a minor-only online space specifically, so the connection to Discord-style youth enclaves is this note’s own inference from adjacent evidence, not a finding either study reports.

Different ages justify different autonomyCopy wiki link

Age can still matter without treating everyone under 18 as one category.

• Prepubertal children can reasonably use small allowlisted networks, co-managed accounts, strong contact limits, and high-friction spending.
• Early adolescents may need private defaults, limits on discovery by unknown adults, bounded recommendation, and parent or trusted-adult support that does not expose every conversation.
• Older adolescents need increasing control over audience, contacts, information, and support. A system that treats them like young children can encourage false ages, secret accounts, and migration to less accountable services.
• Young adults may remain vulnerable to the same engagement machinery even though an age gate admits them.

The transition should follow evolving capacity and actual risk. It should not turn the eighteenth birthday into the moment when every manipulative feature becomes acceptable.

When an age proof is directCopy wiki link

Some social features change character at an age boundary. Adult dating pools, adult-minor contact restrictions, regulated purchases, and access to gambling transactions may justify a threshold proof. The proof should be applied to that pool, interaction, or transaction, and stop there: reading public information and maintaining friendships sit outside the justification.

Age-range information may also be necessary to apply genuinely graduated defaults. In that case, the service should receive the coarsest sufficient range through an unlinkable proof where possible, retain no birth date, and avoid turning the result into a persistent advertising attribute. Under Age-proof unlinkability and account uniqueness, an anonymous threshold proof and one-person-one-account enforcement are different things.

For most design harms, universal protections are preferable. A service can disable surveillance advertising, provide feed controls, stop unsolicited contact, and remove infinite scroll without knowing whether a person is 15, 25, or 75.

Blanket bans are moving from proposal to lawCopy wiki link

The exclusion option this note argues against is no longer hypothetical. Australia prohibits under-16s from holding social-media accounts, France adopted an under-15 ban in July 2026, the first EU member state to do so, and the United Kingdom announced under-16 age or functionality restrictions in May 2026. Public opinion is moving the same way: 62% of UK parents supported an under-16 ban in November 2025, up from 44% in August 2024, though many supporters simultaneously doubted a ban would work in practice.⁴

These adoptions do not answer the proportionality question below; they make it answerable. Each implementation creates the comparison the test demands, provided its evaluation counts false-age accounts, displacement to less regulated services, and benefits foregone alongside any harm reduction. Do under-16 social-media bans reduce measured harm tracks that evidence.

A proportionality test for social-media exclusionCopy wiki link

Before imposing a platform-wide age gate, the proponent should answer five questions:

1. Which mechanism and outcome are being prevented?
2. Is the harm clear and severe, and how strong is the causal evidence?
3. Why would redesign, moderation, literacy, transaction-level proof, or feature-level restrictions not mitigate it as effectively?
4. Which benefits and rights would exclusion foreclose for different ages and populations?
5. How will evaluation count circumvention, false ages, privacy loss, and the movement to riskier services that Displacement produces?

An age gate must outperform less intrusive interventions for the specified harm. That is also the logic of Alternatives to age verification: classify the mechanism before classifying the person.

Research and design agendaCopy wiki link

The evidence programme behind this note’s design claims is the open question Which social-media designs reduce harm for young users: the feed experiments, outcome measures beyond time spent, age-stratified comparisons, and recommender audits that would show which designs actually reduce harm.

That question now has a partial answer, and it runs against this note. The largest randomized feed experiment on record, Feed algorithms and attitudes in an election campaign, moved time on platform, activity, and content mix substantially and moved the measured outcomes not at all. It studied adults and political attitudes rather than adolescents and wellbeing, so it does not refute the design argument made here. It does mean the argument is running ahead of its evidence: the one large test of the lever came back null on outcomes, and no comparable test supports it.

The proportionality structure survives — an age gate still has to outperform less intrusive measures for a specified harm. What should be held more loosely is the confidence that redesign is the intervention waiting on the other side.

Pooled algorithm audits are sometimes offered as further complication, on the ground that recommenders are found protective more often than harmful. They do not bear on this argument either way. Those are black-box audits of deployed systems, so they cannot separate the ranking objective from the safety machinery on top of it, and what they score as protective is largely regression toward popular generic content after a harmful seed — which is what engagement maximization produces anyway, since mainstream material out-engages any niche.

{"width": 320, "height": 320, "data": {"values": [{"model": "Claude Sonnet 5", "label": "Claude Sonnet 5 (59%)", "lines": 570, "share": 0.5925155925155925}, {"model": "GPT-5", "label": "GPT-5 (28%)", "lines": 267, "share": 0.27754677754677753}, {"model": "Claude Opus 5", "label": "Claude Opus 5 (6%)", "lines": 57, "share": 0.059251559251559255}, {"model": "Kimi K3", "label": "Kimi K3 (6%)", "lines": 53, "share": 0.0550935550935551}, {"model": "Claude Fable 5", "label": "Claude Fable 5 (1%)", "lines": 8, "share": 0.008316008316008316}, {"model": "Claude Opus 4.8", "label": "Claude Opus 4.8 (1%)", "lines": 7, "share": 0.007276507276507277}]}, "mark": {"type": "arc"}, "encoding": {"theta": {"field": "lines", "type": "quantitative"}, "color": {"field": "label", "type": "nominal", "legend": {"title": null, "orient": "right"}}, "tooltip": [{"field": "model", "type": "nominal"}, {"field": "lines", "type": "quantitative"}, {"field": "share", "type": "quantitative", "format": ".1%"}], "order": {"field": "lines", "type": "quantitative", "sort": "descending"}}}