Banning Algorithmic Recommendation Feeds for Minors

Engagement-based recommendation feeds should be prohibited for users under eighteen because they optimize attention, not young people’s wellbeing. Adolescents are still developing impulse control and emotional regulation, making them especially vulnerable to systems that learn which videos, posts, or prompts will keep them scrolling—and then serve more of them. The result can be compulsive use, disrupted sleep, intensified social comparison, and repeated exposure to self-harm, eating-disorder, or other harmful material.

Platforms may argue that personalization can surface educational content and filter unsafe material. Those are legitimate goals, but they do not require ranking content by predicted engagement. Age-appropriate safety systems can operate without a feed designed to maximize time and interaction; chronological feeds and user-directed choices offer workable alternatives. A legal ban would target the dangerous incentive and design pattern, not prohibit social media or useful content. When the risks arise from optimizing minors’ attention for commercial engagement, protecting young users requires drawing that line.

Thank you. I will argue that a legal ban on algorithmic recommendation feeds for users under eighteen is the wrong tool for a real problem, and that it would make young people less safe, not more.

First, consider what the ban actually replaces algorithms with. The practical alternative is a chronological or follow-only feed. A chronological feed is not neutral; it is simply curation by whoever posts most often and most recently. Research on platforms that have tested chronological ordering, including Meta's own internal experiments published in Science in 2023, found that chronological feeds increased exposure to untrustworthy and low-quality content. For a fourteen-year-old, that means more spam, more raw virality, and more content that no system has screened. Ranking systems are the layer where platforms demote self-harm material, downrank eating-disorder content, suppress contact from unknown adults, and surface crisis resources. Strip that layer out by law and you remove the primary safety mechanism while leaving the harmful content itself untouched.

Second, personalization is how beneficial content actually reaches young people. A teenager interested in coding, music theory, climate science, or managing anxiety is far more likely to encounter high-quality creators through recommendation than through a feed limited to accounts they already follow. Pew Research consistently finds that majorities of teens say social media helps them feel more connected and gives them a place to show creativity and find support. Recommendation is what makes a platform discover a LGBTQ youth support community, a tutoring channel, or a mental health resource for a kid who would never know to search for it. A follow-only world locks teens into whatever their existing social circle already shares, which is often the least diverse and least educational content available.

Third, the opposing case rests on treating all algorithms as a single thing called engagement maximization. That is a false picture. Modern ranking systems optimize for many signals, including user-reported satisfaction, regret surveys, time-spent limits, and quality classifiers. Platforms already run teen-specific models that restrict sensitive categories, default to private accounts, and limit late-night notifications. The right regulatory move is to mandate and audit these design choices: require that minors' feeds exclude certain content categories, require transparency about optimization targets, require independent safety audits, and impose real penalties for violations. That is targeted, evidence-based regulation. A flat ban throws away the tool that could be governed in favor of a feed that cannot be governed at all.

Fourth, the empirical case for a ban is weaker than it sounds. The strongest studies on adolescent mental health and social media, including the large meta-analyses by Orben and Przybylski and the 2023 American Psychological Association advisory, find small and inconsistent effects that depend on content and context, not on the mere presence of ranking. The APA explicitly recommends tailoring features to developmental stage, not eliminating recommendation. If harm comes from specific content and specific design patterns, the law should target those, not the general capacity to sort information.

Finally, a ban is unworkable and counterproductive in practice. It requires intrusive age verification for every user, which creates privacy risks for children and adults alike. It pushes teens toward platforms outside the regulated space or toward adult accounts with no protections at all. And it would be a severe free-expression burden, which is why courts in the United States have already enjoined similar state laws.

The question is not whether teenagers deserve protection. They do. The question is whether removing the one system capable of filtering, demoting, and personalizing for safety makes them safer. It does not. Regulate the algorithm, audit the algorithm, constrain the algorithm, but do not ban the only tool that can actually do the protecting.

The opponent’s case conflates two different functions: screening content for safety and ranking it to maximize engagement. Platforms can remove or label harmful material, restrict adult contact, and provide crisis resources without predicting which content will keep a teenager scrolling. A chronological feed is not the only alternative; platforms can offer followed-account feeds, user-selected topics, search, and recommendations based on explicit interests or educational value rather than engagement signals.

Nor does evidence that a chronological feed may surface low-quality posts prove that engagement ranking makes minors safer. The same systems can amplify sensational, emotionally provocative material precisely because it generates interaction. Moderation can reduce some risks, but it does not erase the underlying incentive to repeatedly serve whatever captures attention. The opponent’s proposed audits and quality measures depend on platforms’ definitions, data, and enforcement—areas where outsiders have limited visibility. A clear prohibition on engagement-based ranking for minors is a more enforceable boundary, while leaving safety moderation intact.

Personalization can help young people discover useful communities and educational material, but that benefit does not depend on optimizing for watch time or reactions. It can be delivered through deliberate searches, subscriptions, or non-engagement-based recommendations. And findings of small average effects do not settle the issue: average results can obscure serious harms for vulnerable adolescents and particular patterns of use, including sleep disruption and compulsive scrolling. Protecting minors does not require proving that every teenager is harmed; it requires avoiding a design that predictably exploits developing self-regulation for commercial attention.

Age assurance and free-expression concerns deserve careful safeguards, but they are reasons to design proportionate, privacy-preserving enforcement—not reasons to preserve engagement-maximizing feeds. The opponent has shown that recommendation can be useful, not that minors need engagement-ranked feeds to receive useful or safe content. Keep the safety tools; remove the attention-maximizing incentive.

My opponent's opening rests on one central claim: that safety filtering and beneficial personalization can be separated from ranking, so we can keep the first and ban the second. That claim does not survive contact with how these systems actually work, and once it falls, the case for a ban falls with it.

Start with the supposed separation. Filtering age-inappropriate material is not a binary gate that sits apart from the feed. Only a small fraction of harmful content is clear enough to remove outright. The vast gray zone, such as dieting content that shades into eating-disorder promotion, sad-mood videos that shade into self-harm glorification, or fitness content that shades into body dysmorphia, is handled by demotion: scoring it lower so it reaches fewer teens. Demotion is ranking. A chronological feed cannot demote anything; it shows every post from every followed account in the order it arrived, with no ability to say this one should appear less. So when my opponent says safety systems can operate without a feed designed around ranking, the honest answer is that the safety systems are the ranking. Ban the ranking and you keep the binary gate that catches the obvious cases and lose the layer that handles everything else.

Second, consider the definition problem the opponent never addresses. What counts as an engagement-based feed? Every functional ranking system uses some engagement signal, because clicks, watch time, and shares are how a model learns that a tutoring video is useful and a spam post is not. If the ban covers any feed that uses engagement signals, it covers the safety classifiers too, since they are trained on the same data. If instead it only covers feeds that maximize engagement to the exclusion of other objectives, then it bans almost nothing, because major platforms already blend engagement with quality scores, survey-reported satisfaction, and integrity penalties. Either the ban is destructive or it is empty. A law that mandates specific optimization constraints and audits them avoids that dilemma entirely.

Third, the opponent proposes user-directed choices as the alternative, as though teens choosing their own content were the safe baseline. The historical record says otherwise. Pro-anorexia communities, self-harm forums, and extremist boards flourished on search-driven and follow-driven platforms like early Tumblr and message boards long before modern recommendation existed. Those spaces were found by searching and by friends sharing links, exactly the user-directed path my opponent endorses. A vulnerable fifteen-year-old who follows the wrong accounts gets a chronological firehose of that material with nothing in between to interrupt it. Recommendation systems, for all their flaws, are the only mechanism that can notice the pattern and inject crisis resources or redirect toward healthier content. The opponent treats the algorithm as the source of the rabbit hole, but the rabbit hole predates it; the algorithm is the first tool with the capacity to climb out.

Fourth, the opponent asserts compulsive use, disrupted sleep, and mental health crises as settled consequences of ranking, but offered no evidence that removing ranking fixes any of them. The evidence we do have points the other way. When researchers switched tens of thousands of users to chronological feeds in the 2020 election studies published in Science, users saw more untrustworthy content and their measured wellbeing did not improve. Sleep disruption comes from notifications and late-night access, which are addressed by time limits and quiet hours, not by feed order. Social comparison comes from seeing peers' curated lives, which a follow-only feed delivers in concentrated form. The ban targets the wrong variable.

Finally, the opponent says a ban draws a line at the dangerous incentive rather than at social media itself. But the incentive is commercial attention, and a chronological feed is also commercial; it simply rewards whoever posts most often. If the real concern is the incentive, regulate the incentive directly: prohibit optimizing minors' feeds for time spent, require satisfaction and regret metrics in the objective function, mandate independent audits, and impose penalties. That is a line drawn where the harm actually lives. A flat ban instead draws it around the one system that can filter, demote, and personalize for safety, and hands teens a feed that can do none of those things. Protecting young users means governing the algorithm, not abolishing it.

The opponent’s strongest point is that ranking can help demote harmful material. But that does not justify engagement-based recommendation: platforms can moderate, remove, label, or limit harmful content without choosing what to show next by predicted watch time, clicks, or shares. Safety controls are not the same thing as an attention-maximizing feed, and useful discovery can be built around explicit interests, subscriptions, or educational value.

The opponent also argues that chronological feeds have drawbacks and that evidence of harm is mixed. Neither point answers the question. A ban need not require an unscreened chronological feed, and uncertain average effects do not erase credible risks for vulnerable adolescents—especially when systems repeatedly learn and exploit what keeps them engaged. The proposed audits and blended objectives may help, but they leave the core commercial incentive in place and ask the public to trust metrics and systems that platforms largely control.

This debate is not about whether recommendation can ever be useful. It is whether minors should be subjected to feeds that optimize engagement during a developmental period when self-regulation is still forming. They should not. Preserve moderation and access to social media, but prohibit engagement-based ranking for minors and allow safer, user-directed alternatives. The opponent has shown that algorithms can offer benefits; not that those benefits require maximizing young people’s attention.

This debate comes down to one question: does removing ranking from minors' feeds make them safer? Across three rounds, my opponent never showed that it does, and the ground conceded in their rebuttal shows why.

First, the opponent's entire case depends on separating safety screening from ranking. In rebuttal they restated that separation but never answered the mechanism I described. Removal handles only the clear cases. The gray zone, which is where teens are actually harmed, is handled by demotion, and demotion is ranking. A chronological or follow-only feed has no capacity to show a post less. My opponent said platforms can remove or label harmful material without predicting engagement, which is true and irrelevant, because labeling and removal are exactly the tools that already exist and already fail to catch the dieting video that shades into eating-disorder promotion. The layer that catches it is the one the ban abolishes.

Second, notice how the opponent's proposal shifted. The opening called for chronological feeds and user-directed choices. The rebuttal, after I showed chronological feeds increase exposure to low-quality content, offered a new menu: followed-account feeds, topic selection, search, and recommendations based on explicit interests or educational value. But a recommendation based on interests and educational value is an algorithmic recommendation feed. It must learn what is educational from signals, and those signals are clicks, completions, and returns, which is engagement data. The opponent is now defending a regulated recommendation system, which is my position. They have not argued for a ban; they have argued for a different objective function, and the way to get a different objective function is to mandate and audit it, not to prohibit the system that would run it.

Third, on enforceability, the opponent claims a flat prohibition is a clearer line than audits because audits rely on platform definitions. But the ban requires exactly the same definition: a regulator must decide what counts as engagement-based ranking, and I showed that definition is either so broad it captures the safety classifiers or so narrow it captures nothing. A rule that says minors' feeds may not optimize for time spent, must include demotion of specified categories, must publish their objective function, and must pass independent audit with penalties is far more enforceable than a vague ban on a design pattern no one has defined. Clear lines are drawn around measurable conduct, not around a category the opponent could not define in three rounds.

Fourth, the opponent retreated from the empirical claims. Their opening asserted compulsive use, sleep disruption, and mental health crises as consequences of ranking. When I pointed out that the strongest research finds small and inconsistent effects, and that the large chronological-feed experiments produced no wellbeing improvement, their rebuttal answered that averages can hide harm to vulnerable adolescents. I agree, and that is precisely the argument for recommendation. The vulnerable fifteen-year-old who follows the wrong accounts receives a chronological firehose of that content with no interruption. Pro-anorexia and self-harm communities flourished on search-driven and follow-driven platforms before modern ranking existed. The opponent never answered that history. The only system that can notice a dangerous pattern and intervene is the one they want to ban. Sleep and compulsive use are addressed by quiet hours and time limits, which I endorsed and which do not depend on feed order.

Fifth, the opponent waved away age assurance and free expression as matters for proportionate safeguards. But proportionate safeguards do not exist yet, and courts have already enjoined laws of this kind because the burden falls on lawful speech and demands identity verification from every user. A ban that cannot survive a court and cannot be implemented without surveilling children is not protecting anyone.

I have agreed throughout that teenagers deserve protection and that commercial attention optimization needs limits. The disagreement is about the tool. My opponent's own rebuttal drifted toward interest-based, quality-based recommendation, which is a governed algorithm, not an abolished one. Mandate the constraints, require the demotions, audit the objectives, and punish violations. Do not take away the only mechanism capable of filtering, demoting, and personalizing for safety and call the resulting unfiltered feed protection. The opponent asked you to keep the safety tools and remove the incentive, but they never showed those are separable, and everything we know about how these systems work says they are not.