Field Notes

question

Which social-media designs reduce harm for young users?

StatusWorkingUpdatedStands on3 sourcesSupportsnotes

Which concrete design changes — feed construction, defaults, friction, and governance — measurably reduce harm for children and adolescents, and which merely relocate it?

Why it matters

Less harmful social media argues that harm follows from engagement-maximizing design rather than from access itself, and that an age gate must outperform less intrusive interventions for a specified harm before exclusion is justified. Alternatives to age verification makes the same move: classify the mechanism before classifying the person. Both arguments stand on design-level evidence that is currently thinner than the policy debate assumes. Answering this question decides whether safer-design claims can carry the weight those notes place on them, and it directly informs EU age verification and Internet privacy.

What is already known

Restriction experiments and the Commission’s preliminary findings against major platforms establish that engagement design is a regulable object, not a fixed background condition. Existing evidence rarely separates design effects from time-spent effects, rarely stratifies by developmental stage, and rarely measures benefits foregone alongside harms avoided.

Demand-side evidence exists where outcome evidence does not. The Internet Matters Online Safety Act report found United Kingdom children welcoming restrictions on stranger contact (77%) and limits on features such as livestreams (74%), wanting limits on engagement mechanics such as streaks and location sharing, and naming time online, not stranger contact, as their most immediate day-to-day concern. These are preferences reported by users, not measured design effects, but they identify which interventions have a constituency among the people they constrain.

None, on current evidence — and the best test of the lever came back null

The honest answer is that no design change has been shown to reduce harm for young users, and the reason is worse for the vault’s argument than a simple absence of studies would be.

The largest randomized test of feed construction that exists, Feed algorithms and attitudes in an election campaign, switched Facebook and Instagram users to a reverse-chronological feed for three months. The design change worked as a design change: time on platform and user activity fell substantially, and the content mix shifted measurably in several directions at once — more political and untrustworthy material, less incivility, more moderate sources on Facebook. The downstream attitudes did not move. Issue polarization, affective polarization, political knowledge, and other key attitudes showed no significant change.

Algorithm audits are often read as pointing the same way, and the reading needs care. Hilbert and colleagues pooled 151 audits from 33 studies across YouTube, Search, Twitter, Facebook, TikTok, and others, finding roughly 8 to 10 percent of recommendations “bad” and, in the subset reporting it, about a quarter “good.”1 Their headline is that algorithms pull users out of a self-inflicted rabbit hole more often than into one: 40 percent good-to-good transitions against 11 percent bad-to-bad.

What was coded as “good” also matters, because it is not crisis resourcing. The high-scoring audits are accurate influenza-vaccine pages in Google results, vaccine-related YouTube content, and age-appropriate material surfaced on children’s channels — utility and accuracy, not self-harm steering. The “protective” pattern is a transition statistic: after a bad seed, most subsequent recommendations are not bad.

The most economical reading of that is regression toward popular generic content, which an engagement-maximizing recommender produces by default, because mainstream material has higher expected engagement across a broad audience than any niche does. On that reading the audits are consistent with engagement maximization rather than evidence against it, and they neither support nor undermine the design argument this question is testing.

What this does and does not license

Three limits keep this from settling the question outright. The feed experiment studied adults, measured political outcomes rather than wellbeing, and ran for three months. It is evidence about the general proposition that feed construction drives outcomes, not about adolescent harm specifically, and the adolescent design-experiment base remains essentially empty.

What it does establish is that the premise Less harmful social media and Alternatives to age verification lean on — that harm follows from engagement-maximizing design, so redesign is the more direct remedy — is not currently supported by outcome evidence, and has one large well-powered result against it.

This is not an argument for age gates. If design changes do not move outcomes, that is evidence about the harm model itself rather than a reason to prefer exclusion, and an age gate would still have to show the effect that redesign failed to show. The proportionality test in those notes survives intact. What weakens is the confidence that a better-designed feed is the thing on the other side of it.

What would settle it

  • Bounded and user-directed feeds tested against engagement-maximizing feeds in real social networks.
  • Sleep, unwanted exposure, harassment, regret, connection, support, and expression measured alongside time spent.
  • Effects studied separately for prepubertal children, early adolescents, older adolescents, and young adults.
  • Universal safer defaults compared with age-tailored defaults that require anonymous age-range proof.
  • Young people given meaningful power in study design, product governance, and interpretation of results.
  • Audits of whether platform safety tools alter the core recommender system or merely add controls that engagement design can override.

  1. Hilbert, Thakur, Flores, Zhang, and Bhan, “8–10% of algorithmic recommendations are ‘bad’, but… an exploratory risk-utility meta-analysis and its regulatory implications”, International Journal of Information Management (2023), pooling 151 algorithmic audits from 33 studies. ↩

Built on 3 sources (3 external).

Working out connections…