A Minimum Age Online

Bar under-13s from social media accounts and switch off algorithmic feeds for everyone under 17.

AI evaluation · not yet reviewed by a human

This evaluation was produced and sourced by an AI model; a human review is still pending. Figures and conclusions may still change. The review log is at the foot of the page.How review works →

Platforms already forbid accounts for under-13s in their terms of service and do almost nothing to enforce it. The bill before Congress would make the age limit a legal duty, require platforms to take reasonable steps to know who is under it, and separately switch off algorithmically personalised recommendations for every user under 17, leaving them a feed of what they chose to follow. Australia went further in December 2025, barring under-16s from ten named platforms, and removed 4.7 million accounts in the first week. This evaluation compares five years under the American bill against five years without it.

Balance

Balanced · 0.56 previous scale

Balance on the previous scale. The Bilanz 2.0 simulation is not yet available for this evaluation. The category comes from the share of the debate on the pro side (r).

For 43 · 56 % Against 35 · 44 %

Size class: medium Scale of this evaluation: Normalised Impact — unitless, calibrated to this topic. For comparison: one point here is worth roughly 200 million euro per year. The two largest figures here are the same effect measured in opposite directions on the same 17 million teenagers: what a feed costs them and what it gives them. The first is set at the size found when Facebook arrived at American colleges, the second at two fifths of it. Nobody has measured the second, and the balance depends on it. How we score →

Arguments for

Arguments against

5 arguments evaluated · Scoring v1.3 Δ absolute +8

Arguments — For

2 arguments

Adolescents in better mental health

32of 100

The best causal evidence on this question comes from Facebook's arrival at American colleges one at a time between 2004 and 2006. Where it arrived, symptoms of depression and anxiety rose, and the mechanism the authors identify is unfavourable comparison with other people. The bill removes the part of the product that does the comparing.

Value 9 · HealthImpact 8Plausibility 4.5

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is the mental health of people between about eleven and seventeen: how much they sleep, whether they are anxious in a way that interferes with school, how they feel about their own life against everyone else's. This site places it in the class it uses for life and health, one step below the top because adolescent low mood is generally recoverable and because what is measured is a symptom index rather than a diagnosis. The weight does not rise because the people concerned are young; what youth changes is how long the consequences run, and that belongs to the Impact. The academic consequences are counted separately in the next argument. The value sits one step below the maximum: the stream is health, and health that can be regained.

About 17 million Americans are aged 13 to 16, and several million more are under 13 and hold accounts anyway. The size of what a feed does to them comes from the one clean measurement available: Facebook arrived at American colleges at different dates, and where it arrived, symptoms of poor mental health rose by about a tenth of a standard deviation — roughly a fifth of the effect the same authors compute for losing one's job ^[4]. Removing algorithmic recommendation is not the same as removing the platform, so 30 percent of that effect is used here, in a range from 10 to 60 percent: about 0.025 standard deviations across 17 million teenagers, plus a larger effect on the roughly 4 million under-13s who would lose accounts entirely. A standard deviation on such an index is worth roughly 0.06 quality-adjusted years, in a range from 0.03 to 0.12. The two together give about 40,000 quality-adjusted years a year. The Impact is the largest in this debate, and it is large because of how many adolescents there are rather than how much happens to each of them.

▸ Show calculation ▾ Hide calculation

	Americans aged 13 to 16		17 million people
×	Improvement in symptoms of poor mental health Setting, range 10 to 60 percent of the measured effect: Facebook's arrival at a college raised symptoms by about a tenth of a standard deviation, and this measure removes one feature rather than the platform [4]	0.025 standard deviations	425,000 standard-deviation-people
+	Under-13s who lose accounts entirely a larger effect, because the account goes rather than the feed	4 million × 0.06 standard deviations	665,000 standard-deviation-people
×	Quality-adjusted years per standard deviation Setting, range 0.03 to 0.12	0.06	39,900 quality-adjusted years
×	Value of the years the value of a healthy life year used across this site	40,000 euro each	1,596 million euro
÷	Normalised Impact scale of this evaluation	200 million euro a point	8

Score 8 Impact × 9 Value × 4.5 Plausibility ÷ 10 = 32 of 100

This is the most contested empirical question in the area and the evidence divides by design rather than by conclusion. The counterfactual in the study carrying this figure is colleges that had not yet received Facebook, compared against those that had, on the same survey instrument before and after ^[4]. The confounder that would otherwise dominate — that early colleges differ from late ones — is what the staggered rollout absorbs, and reverse causation cannot arise, since a student's mood did not determine when Facebook expanded. Against that, meta-analyses of the wider literature repeatedly find effects near zero, and they are right to: most of that literature is cross-sectional and measures who uses social media rather than what it does. Two things limit the transfer here. The study is about college students in 2004 and this measure concerns adolescents in 2026, whose product is a video recommendation engine rather than a list of friends' profiles. And the study measures a platform arriving, while this measure removes one feature of it — which is why only 30 percent of the effect is carried across. The Plausibility is below the middle: the design behind the number is sound, and almost everything about the transfer to today's product is assumed.

evidence basis: Study · P ceiling 8 identification: Quasi-experimental · rung ceiling 8 band: Chain closed, unevidenced · P 4–5

Counterfactual: American colleges that had not yet received Facebook, on the same survey instrument. Design: quasi-experimental — generalised difference-in-differences on a staggered rollout (Braghieri, Levy and Makarin, American Economic Review 2022 [4]). Confounder: early-adopting colleges differing from late ones, absorbed by the staggered timing. Direction: no reverse causation, student mood did not determine Facebook's expansion order. Ceiling: quasi-experimental 8.0 binds below the studie ceiling of 9.0; a context transfer of 3.5 applies for two separate gaps — college students in 2004 against adolescents in 2026, and a platform arriving against one feature being switched off. The size doubt sits in the 10 to 60 percent band, not in P.

The chain is named and its one measurement is good: a staggered rollout, a tested comparison, no reverse causation. What is missing is a measurement on today's product. The meta-analyses that find effects near zero are not a counter-finding on the same outcome, because they are cross-sectional and measure who uses social media rather than what it does — a different question under a design that cannot answer this one. Read back: about half the time, switching off algorithmic feeds improves adolescent mental health by roughly the amount assumed here.

Open: Australia removed 4.7 million under-16 accounts on a single date in December 2025. A comparison of adolescent mental health measures against a similar country over three years would replace the whole transfer with a direct measurement and could carry P to 7.

Better years at school

11of 100

The same study that found worse mental health found students reporting that it interfered with their academic work. That is a second consequence rather than the same one told twice: what a person learns between thirteen and seventeen shows up in their earnings for forty years.

Value 8 · Life chancesImpact 3.4Plausibility 4

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is what a young person learns and what that is worth to them later: qualifications gained, courses not dropped, the difference between finishing and not. This site places it in the class it uses for subsistence and life chances, above economic output and below life and health. It is genuinely separate from the mental health counted above — a student can be unhappy and still learn, or content and still fall behind — and it is counted only to the extent the evidence links the two. What the schooling is worth to the wider economy rather than to the person is not counted, because it would be the same hours twice. The value sits in the upper part of the scale, at the level this site uses for subsistence and life chances.

About 17 million Americans are of secondary school age. The study behind the previous argument found that students at colleges receiving Facebook were more likely to report that poor mental health had impaired their academic performance, which is the link this argument uses ^[4]. Translating that into learning is a construction: a one percent improvement in what a student takes from a school year is assumed, in a range from a third of a percent to three percent. A year of American schooling is worth roughly eight percent on lifetime earnings, so one percent of a year is worth about 40 euro of present value per student, in a range from 15 to 120. Across 17 million students that is 680 million euro a year. The figure counts the flow of one cohort-year at a time rather than the accumulated stock, which keeps it comparable to the other annual figures here. The Impact is under half the mental health gain, which is the right order given that the academic finding in the underlying study is a self-report rather than a grade.

▸ Show calculation ▾ Hide calculation

	Americans of secondary school age		17 million students
×	Improvement in what a student takes from a school year Setting, range 0.33 to 3 percent: the underlying study finds more students reporting that poor mental health impaired their academic work [4]	1 %	170,000 student-year-percent
×	Present value of that to each student Setting, range 15 to 120 euro: a year of American schooling is worth roughly eight percent on lifetime earnings, and this is one percent of a year	40 euro	680 million euro
÷	Normalised Impact scale of this evaluation	200 million euro a point	3.4

Score 3.4 Impact × 8 Value × 4 Plausibility ÷ 10 = 11 of 100

The link is real in the source and thin in the application. The counterfactual is the same students at colleges without Facebook, and the academic finding rests on the same staggered rollout that carries the mental health result, so its identification is equally good ^[4]. What is much weaker is everything after it. The study measures students saying their work suffered, not grades or completion, and self-reported impairment and actual learning are not the same quantity. The step from there to lifetime earnings runs through a return-to-schooling figure estimated on whole years of education rather than on marginal quality within a year, which is a different object. The counter-mechanism is unaddressed: a student who spends less time on a feed does not necessarily spend it studying, and the American time-use evidence on what displaces what is not settled. Reverse causation does not arise in the underlying design. The Plausibility is below the middle: the finding it starts from is identified and every step after it is assumed.

evidence basis: Study · P ceiling 8 identification: Quasi-experimental · rung ceiling 8 band: Chain closed, unevidenced · P 4–5

Counterfactual: American colleges without Facebook, same staggered rollout as pro-1 [4]. Design: quasi-experimental for the academic-impairment finding; everything after it is construction. Confounder: displaced time not going to study, unaddressed. Direction: no reverse causation in the underlying design. Ceiling: quasi-experimental 8.0 binds; a context transfer of 4.0 applies because the source measures self-reported impairment among college students and this argument prices learning among school pupils, valued with a return-to-schooling figure estimated on whole years. Band: chain closed but unevidenced — the chain is named and the displacement counter-mechanism is stated; only the measurement is missing.

Nothing measured argues against the claim; what is missing is any measurement of learning rather than of self-reported impairment. The counter-mechanism — that time off a feed does not become study time — is named and unresolved. Read back: about half the time, removing the feed improves a school year by roughly the amount assumed here.

Open: Australia's ban applies to a whole cohort at one date. Comparing school outcomes for Australian 15-year-olds against a comparable country over three years would measure this directly and could carry P to 6.

Arguments — Against

3 arguments

Everyone has to prove who they are

21of 100

A platform cannot know who is under 13 without checking everyone. That means an identity document, a face scan or a third-party check for roughly 250 million American adults, and a new record connecting a legal identity to an account that did not previously require one.

Value 9 · Basic rightsImpact 4.2Plausibility 5.5

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is two things that arrive together: the time everyone spends proving their age, and the end of anonymous participation in public discussion. This site places the second in the class it uses for the constitutional core and treats the first as time under compulsion, which sits near the top of the scale as well; the combination is priced just below life and health. The point is not that identity checks are wrong — banks do them daily — but that a social platform is where a great deal of political speech now happens, and a system that connects every account to a document changes what can be said on it. Nothing here is counted for the risk of that database leaking, which is real and unquantified. The value is high because the stream is anonymity in public speech alongside hours of compelled time, not a convenience.

Roughly 250 million Americans hold at least one social media account. Verifying them, and re-verifying when a check expires or a device changes, is put at 12 minutes a person a year, in a range from 4 to 30 — uploading a document, waiting, retrying when a face scan fails. That is 50 million hours a year, valued at the rate this site uses for time a person must spend with nothing in return. The larger part is the loss of anonymity itself, which has no market price: 2 euro per account holder per year is used, in a range from 0.5 to 10, which is the sort of figure surveys of willingness to pay for privacy return and which is set here rather than found. Together that is about 842 million euro a year. The United Kingdom's experience since 2025 suggests the friction is real: verification failure rates on first attempt run well above what the providers projected. The Impact is half the mental health gain, which is what makes this debate close rather than clear.

▸ Show calculation ▾ Hide calculation

	Americans holding at least one social media account		250 million people
×	Time verifying and re-verifying Setting, range 4 to 30 minutes: uploading a document, waiting, retrying when a face scan fails	12 minutes a year	50 million hours
×	Value of forced time the rate this site uses for time a person must spend with nothing in return	6.85 euro an hour	342 million euro
+	Value set on the end of anonymous participation Setting, range 0.5 to 10 euro a person: the order of magnitude surveys of willingness to pay for privacy return; set here rather than found	250 million × 2 euro	842 million euro
÷	Normalised Impact scale of this evaluation	200 million euro a point	4.21

Score 4.21 Impact × 9 Value × 5.5 Plausibility ÷ 10 = 21 of 100

That checking everyone's age costs everyone time is not a prediction, and the mechanism has no behavioural step in it: the law requires reasonable steps, reasonable steps require a check, a check takes time. The counterfactual is the current position, in which nobody verifies anything. What is estimated is how long a check takes at scale, and here there is now real experience rather than projection — the United Kingdom has required age assurance since 2025 and its providers publish completion and failure rates ^[1]. The confounder that would matter is technology: on-device age estimation is improving quickly and could make most checks invisible within the horizon, which would cut this figure substantially and is unresolved. The privacy half of the argument is a stated price, not a measurement, and it carries most of the quantity — that is the weakness here rather than the mechanism. Reverse causation does not arise. The Plausibility is at the upper end of what a projection can carry: the time cost is close to arithmetic and the price on anonymity is set rather than found.

evidence basis: Projection · P ceiling 6 identification: Definitional · no rung ceiling

Counterfactual: the current position, in which platforms verify nobody. Design: definitional — a legal duty to know a user's age requires a check, and a check takes time; no behavioural link carries the quantity. Confounder: on-device age estimation improving enough to make checks invisible, which would cut the figure; unresolved. Direction: not applicable. Ceiling: projektion 6.0 binds; P sits below it because the privacy half of the quantity is a stated price rather than a measurement.

Teenagers lose what they get from it

12of 100

The same feed that produces unfavourable comparison also produces the group chat, the club, the person who has the same rare illness, and for teenagers who are isolated where they live, the only people who are like them. Removing it removes both. Nobody has measured the second half.

Value 9 · HealthImpact 3.2Plausibility 4

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is the same one the first argument counts, running the other way and on the same people: connection, belonging and support, and their absence. It carries the same weight for the same reason — what is at stake is adolescent mental health, one step below the top of the scale because it is recoverable. Nothing here is counted for what parents gain or lose, which is a different stream on different people and is not priced anywhere in this evaluation. Counting it separately rather than netting it against the gain is deliberate: the two effects fall on overlapping but not identical groups, and a teenager who is isolated where they live is not the same teenager who is worn down by comparison. The value sits one step below the maximum: the stream is health, and health that can be regained.

The same 17 million adolescents are affected, and the effect runs the other way. Its size is set at two fifths of the harm counted in the first argument, in a range from a fifth to the full amount — which is to say, this evaluation assumes the feed is on balance bad for teenagers but not overwhelmingly so. That is a judgment and it is the number this whole debate turns on. It is not arbitrary: the measured net effect of Facebook's arrival on college students was negative, which means the benefits were real and smaller than the harms, and two fifths is a reading of how much smaller. The loss falls hardest on adolescents who are isolated offline — those in rural areas, those with rare conditions, those whose identity is unwelcome where they live — for whom the feed is not one social channel among many. That concentration is not priced separately, which understates this argument. The Impact is two fifths of the gain it offsets, by construction, and there is no measurement anywhere that would put it somewhere else.

▸ Show calculation ▾ Hide calculation

	Mental health gain counted in the first argument [4]		1,596 million euro
×	Share of it that is connection lost rather than comparison spared Setting, range 20 to 100 percent: the measured net effect of a feed arriving was negative, which bounds this from above; how far below is not measured anywhere	40 %	638 million euro
÷	Normalised Impact scale of this evaluation	200 million euro a point	3.19

Score 3.19 Impact × 9 Value × 4 Plausibility ÷ 10 = 12 of 100

This argument is grounded in the same study as the first one and reads it differently. The counterfactual is identical: colleges before and after Facebook arrived ^[4]. What that study measures is a net effect, so the benefit side is inside it rather than separate from it, and no design isolates the two components — which is why the figure here is a share of the other rather than a measurement of its own. The confounder in the qualitative literature that does examine the benefits is severe: teenagers who say social media helps them are describing their own experience, and the ones most likely to say so are the ones most attached to the product. Reverse causation is unresolvable there. What can be said is that the net effect being negative bounds this argument from above — the benefits cannot exceed the harms, or the measured sign would flip. The Plausibility is below the middle: the stream certainly exists, its size is derived from the other side of the ledger rather than observed, and no study separates the two.

evidence basis: Mechanism · P ceiling 6 identification: Mechanistic · rung ceiling 6 band: Chain closed, unevidenced · P 4–5

Counterfactual: the same colleges before and after Facebook arrived — the same source as pro-1 [4], which measures a net effect and therefore cannot separate benefit from harm. Design: mechanistic — the benefit component is inferred as a share of the net rather than identified. Confounder: self-reported benefit correlating with attachment to the product; unresolvable in the qualitative literature. Direction: reverse causation is live for the self-report evidence. Ceiling: mechanistic 6.0 binds. Band: chain closed but unevidenced — the chain is named, the net-effect sign bounds the argument from above, and only the split has never been measured.

Nothing measured argues against the claim; what is missing is any design separating the benefit of a feed from its harm. The net effect being negative bounds this argument from above. Read back: about half the time, the connection teenagers lose is worth roughly two fifths of the harm they are spared.

Open: Australia removed 4.7 million under-16 accounts on one date. A comparison of loneliness and support measures for Australian 14-year-olds against a comparable country would give the first direct reading of this side of the ledger.

Small sites close rather than comply

2.4of 100

Age assurance costs money per user and a hobby forum has no users to spread it over. When the United Kingdom imposed a similar duty, a number of small British forums shut down rather than pay for it. The platforms the bill is aimed at can afford it easily.

Value 7 · Public debateImpact 0.8Plausibility 4.5

▸ Show reasoning & sources ▾ Hide reasoning & sources

The stream is places for people to talk to each other that stop existing: a forum for a rare disease, a local history board, a community that never had a business model. This site places that in the class it uses for participation and the conditions of public debate, above money and below health. What is priced is the loss of the venue rather than any particular conversation, and no attempt is made to weigh a small forum against a large platform by size — the argument is about the ones for whom no substitute exists. The compliance cost to large platforms is not counted here, because they can carry it and it changes nothing about what they offer. The value sits in the upper middle of the scale, at the level this site uses for the conditions of public debate.

Age assurance is priced per check, which makes it nearly free for a platform with a billion users and prohibitive for a forum with four thousand. The British experience since 2025 is the closest available: a number of small forums closed rather than implement the checks their online safety duties required, and others geoblocked the country ^[1]. The American population is five times larger and its small-site ecosystem correspondingly so; 150 million euro a year is used for the value of what closes, in a range from 30 to 500 million. That is a price set rather than derived, and it deliberately excludes the compliance spending of large platforms, which is a cost to them and not a loss to anyone else. Whether the bill's wording would reach small forums at all depends on how it defines a covered platform, which is not yet settled. The Impact is the smallest in this debate and it is the one that a narrower definition of a covered platform would remove entirely.

▸ Show calculation ▾ Hide calculation

	Value of small forums and community sites that close rather than verify Setting, range 30 to 500 million euro: scaled from British closures after equivalent duties took effect; excludes large platforms' compliance spending, which is a cost to them rather than a loss to anyone else [1]		150 million euro
÷	Normalised Impact scale of this evaluation	200 million euro a point	0.75

Score 0.75 Impact × 7 Value × 4.5 Plausibility ÷ 10 = 2.4 of 100

The mechanism has a documented precedent and no measurement. British forums did close after age assurance duties took effect, which establishes that the effect is real rather than theoretical ^[1]. The counterfactual there is the same forums before the duty, which is weak — several were already struggling and the duty was the occasion rather than the cause for some of them, and nobody has separated the two. The chain is short and visible: a per-user cost falls on a site with no revenue per user, and the site closes. The counter-mechanism is strong and only partly answered: the American bill is aimed at platforms with algorithmic feeds and may simply not cover a message board, in which case this argument is worth nothing. That is a drafting question rather than an empirical one. Reverse causation is a live concern for the British closures. The Plausibility is below the middle: the effect is documented, its size is a stated price, and whether it applies at all depends on wording that does not exist yet.

evidence basis: Precedent · P ceiling 6 identification: Mechanistic · rung ceiling 6 band: Chain closed, unevidenced · P 4–5

Counterfactual: the same British forums before the age assurance duty — weak, since several were already struggling. Design: mechanistic — a documented precedent without a design separating the duty from prior decline [1]. Confounder: forums closing for unrelated reasons and attributing it to the duty; unresolved, and reverse causation is live. Direction: contested for the British closures. Ceiling: mechanistic 6.0 binds below the praezedenz ceiling of 8.5. Band: chain closed but unevidenced — links named, the drafting counter-mechanism stated, only the price unmeasured.

British forums did close after equivalent duties took effect, so the direction is documented; what is missing is any measure of what they were worth. The counter-mechanism — that the American bill may not cover message boards at all — is a drafting question and is unresolved. Read back: about half the time, a duty of this kind closes small venues worth roughly the amount assumed here.

Open: The final definition of a covered platform in the bill text settles most of this outright, and British forum closure counts since 2025 are already documented and could be scaled.

Summary

Almost everything in this debate is one number seen from two sides: what an algorithmic feed does to seventeen million adolescents. The best evidence for the harm is genuinely good — Facebook arrived at American colleges one at a time, and where it arrived, mental health measurably worsened — but it is twenty years old, it concerns a different product, and it measures a platform appearing rather than a feature being switched off. The benefit side has no comparable measurement at all, because the same study reports a net effect and nothing separates the connection a feed provides from the comparison it invites. This evaluation sets that benefit at two fifths of the harm, which is a judgment rather than a finding, and the balance would tip either way at a third or a half. What can be said without any of that is narrower: enforcing an age limit means checking everyone's age, and a country that does it acquires a record connecting every account to a legal identity, which is a real price and is being paid for a result nobody has yet measured.

Outlook — effect over time

Balanced · 0.56 previous scale

Sources

Route Fifty: How would proposed age restrictions on social media use actually work?. route-fifty.com
Senator Brian Schatz: Kids Off Social Media Act. schatz.senate.gov
eSafety Commissioner, Australia: Social media age restrictions. esafety.gov.au
Braghieri, Levy and Makarin, American Economic Review 112(11): Social Media and Mental Health. aeaweb.org
Pew Research Center: Teens, Social Media and AI Chatbots 2025. pewresearch.org

Last reviewed by Claude Opus 5 · September 6, 2026 · 3× AI, not yet reviewed by a human

September 6, 2026AI reviewClaude Opus 5record updated
i_spanne an allen 5 Argumenten aus den englischen Ketten, kein Transfer im Record. Kategorie kippt von Ausgeglichen (r 0,56) auf Schlechter (P(D>0) 0,15).
September 6, 2026AI reviewClaude Opus 5record updated
i_spanne an allen 5 Argumenten aus den englischen Ketten, kein Transfer im Record. Kategorie kippt von Ausgeglichen (r 0,56) auf Schlechter (P(D>0) 0,15).
September 6, 2026AI reviewClaude Opus 5First evaluation
Created for the English side: the harm is taken from the staggered Facebook college rollout, the benefit is booked as a share of it because nothing measures it.

Evaluations are produced with AI support and reviewed on a schedule for new developments; human passes are marked separately.How we review →