Ban or design?

The regulation setting an age limit for social media was published in the Official Gazette on 10 October 2026 and takes effect on 1 November.^[1] The debate has largely squeezed into one question: is a ban for under-15s right or wrong? This piece shifts the question a little.

It looks less at the ban itself than at the rules written beside it and how they will be measured. We also had the strongest objection to the view most models agreed on written by a separate counter-voice; that objection is in this piece too.

What was introduced

The regulation has two layers. The first is the age limit: children under 15 cannot open a social media account. The second is rules for those above the limit: accounts of 15–18-year-olds will run with the default settings listed in Article 8 of the regulation.

How age is proven is also described. The user presents the platform with an "age token," generated through e-Devlet (the national e-government portal) and carrying only the age threshold. Technical details of the token will be set by a guide the Presidency will publish (art. 12). That guide has not yet been published.

The text also contains parental tools (art. 9) and an obligation to publish annual statistics (art. 10).

Three approaches worldwide

The paths other countries have tried fall roughly into three models. To understand Türkiye's regulation, first look at what they produced.

The line is fading: from design to ban

In the past year the two models have drawn closer. In November 2025 the European Parliament, in a non-legislative resolution, called together for an EU-wide minimum age of 16 (13–16 with parental consent) and a ban on the most harmful addictive design practices.^[19] The country that has tried design rules longest is the United Kingdom.

According to Ofcom research released in March 2026, 72% of 8–12-year-olds there accessed sites and apps with a 13+ limit.^[20] In June 2026 the government announced a ban on the service for under-16s; the rules are expected to take effect in spring 2027.^[18] Türkiye building both models together fits this trend.

How well does a ban hold: Australia's first three months

The most recent measured example of the ban model is Australia. The country's online safety regulator eSafety published findings on the ban's first three months in July 2026.^[2][3] The picture: the ban reduces the number of accounts, but use only a little.

Children using at least one restricted platform

Before85.9%

Month 381.5%

Children with an account

Before52.4%

Month 342.1%

Account ownership fell markedly; use fell only about 4–5 points.^[2][4] The reason is less children's cunning than platforms' enforcement: of children who still have an account, 50.2% say the platform never asked their age and 18% say their age was wrongly estimated.^[3][4] Three months may also be too short to show the ban's main effect: age limits leave their mark mostly on the cohort that has no account yet. That objection is in Thesis 2.

Time limits: Korea's lesson

South Korea introduced a night-time gaming ban in 2011. It was found ineffective^[12] and repealed in 2021.^[13] The shared view of the models in the round is that hour and time bans are the weakest of the three. The counter-voice objected with Norway's school phone ban; both views are in Thesis 5.

An age limit decides who gets in; design rules decide what happens inside. If neither is measured, we cannot know which one works.

Theses

The round's main output is six theses. Each card has the thesis's reasoning, the strongest opposing view and the measure that would change our mind. Below each card is every model's stance on that thesis: defending, partly agreeing, opposing and not addressed. Stances were measured by quotation from each model's own answer.

What Türkiye's model does well

Türkiye did not pick one of these three paths. It is building an age limit, design rules and a central age token produced by the state all at once; a rare combination worldwide.

The least-discussed part publicly is Article 8; why these rules do more work than the age limit is in Thesis 1. For accounts of ages 15–18 the regulation says:

• No swipe or auto-advancing visual-content interface may be offered. This means a ban on infinite feeds and autoplay.
• The account can't be personalised by the child's behaviour; profiling and personalised ads are banned.
• Live streaming can't be opened.
• Screenshots can't be taken.

These articles make Türkiye's 15–18 rules among the strictest in the world. A note: in the round's first summary we too had skipped these four rules, and some models suggested them as "things to do." Yet they are already in the text. What is missing is not the rule itself but being able to see from outside whether the rule is obeyed.

The infrastructure is also ready. So Türkiye can do in three weeks what France couldn't in three years. For the same reason, the risk is also large.

Where it breaks

The models didn't openly oppose any thesis. Where they diverged is partial, but those are the places that decide. The reasoning and measure for the breakpoints are in the thesis cards; here are only where the models diverged and where we ended up.

The token: privacy is decided by operation, not content

Open questions and the measure are in Thesis 3. The divergence here is small but instructive. Studio agent 3 said infrastructure produced by a single institution is the structure most prone to expand over time to thresholds beyond age.

Gemini 3.8 Flash too wrote of the risk of an access map forming on the state side, showing who enters which platform. Gemini 3.1 Pro, by contrast, found the state guarantee positive in that less data goes to platforms.

Where we ended upFix the token's form at two bits (15+, 18+) and publish the specification openly.

Handover and migration: the token changes hands, harm relocates

Five voices (the studio's three agents, Gemini 3.1 Pro and the Gemini app) saw the main leak at home and in out-of-scope apps. Three voices (Claude Opus 5.5, Gemini 3.8 Flash and Grok) counted VPNs with equal weight (Thesis 4).

The real divergence is in the remedy. Gemini 3.1 Pro wanted more identity: NFC ID reading, live face verification. Studio agent 3 argued the opposite: a token without identity passes from hand to hand by nature. The regulation chose privacy, and chose rightly, but it must measure the cost.

Where we ended upThe second view. Handover is managed by measurement, not by prohibition.

The excluded child

15–17-year-olds who can't get a token are the subject of Thesis 6; children whose parents are uninvolved should be added to the groups on the card. Claude Opus 5.5, Grok and studio agent 3 raised this group; the other voices took no stance.

And the ban itself?

Here too there is no single view. Gemini 3.8 Flash argued a blanket ban doesn't protect children, leaving them more vulnerable by pushing them behind adult IDs. Gemini 3.1 Pro and the Gemini app framed the regulation positively, as rebuilding the public sphere on the state's trust infrastructure. Claude Opus 5.5 combined the two: this is an infrastructure decision; both power and risk.

Where we ended upThe ban is defensible. But success is not measured by the ban itself. The ban's strongest defence is in the next section.

The counter-voice: the strongest defence of the ban

Most models met on "what protects is design, not the ban." To see how much of that agreement is real and how much comes from the shape of our question, we gave a separate model (Claude Opus 5.5) one task: write the strongest defence of the ban, with sources. The defence is built in four steps.

1. The problem is not individual children but everyone being there. The cost to a child of not being on a platform grows with their friends being there. Experiments with university students on TikTok and Instagram showed that even active users may prefer the platform not to exist at all; the authors call this a "product-market trap."^[27] Ban advocates call the same situation a "collective action trap": every exception allowing a younger age with parental consent traps everyone again.^[28] This is a coordination problem, and a coordination problem is solved by a rule applied to everyone at once. A design rule leaves the child inside the network.
2. Law changes norms, and norms move before behaviour. The sentence "not before 15" gives parents a non-negotiable footing. Even the 21 drinking age, which everyone knows is circumvented, clearly reduced consumption and the main harms in the US.^[22] The tobacco-21 limit also markedly lowered regular smoking among 18–20-year-olds who had tried smoking; the effect was larger for those whose closest friends smoked at 16.^[29] In Australia in the ban's third month, among 10–15-year-olds those saying "most of my peers use social media" fell from 85.1% to 74%, and those saying "someone without an account is missing out" from 43.3% to 36.3%. Among parents, those saying "it's normal for my child's peers to have accounts" fell from 73.4% to 60%. The report adds a warning: perception can change faster than behaviour, and in month three the accompanying behaviour change is still limited.^[23] The age limit's effect is expected to show most in the 10–12 cohort that hasn't started yet.
3. A one-question rule is the cheapest to enforce. Those who must implement design rules are the party whose revenue comes from the child's attention. As rules multiply, each article can be watered down separately. "Is this account under 15" is a single, clear question.
4. The dependency runs the other way. Article 8 applies only if the user's age is known. So the design layer rests on age infrastructure: the ban works without design rules, design rules don't work without age information.

The counter-voice also wrote its own weak points. The collective-trap evidence comes from adult students, not children. The alcohol analogy is lame too: alcohol's point of sale is physical and enforceable, an account's is not. Most important, in Australia the share of parents of children who said they used social media in the last four weeks who don't know it rose from 23.3% to 33.3%.^[4] The ban may be turning visible use into invisible use.

What this objection taught us

• We split the question in two. "Ban or design?" set the two up as rivals. Yet one is the other's condition: design rules don't work unless age is known. The right question should have been not "which" but "which carries which."
• We didn't weigh the evidence equally. We looked at the ban by its outcome: use after three months. We looked at design rules by their input: 128 changes on platforms. Applying the same measure to both would strain design rules too.^[20][21]
• No model openly opposed. We measured every model's stance on every thesis by quotation from its answer. In 48 pairings there are 33 agreements, 7 partial agreements and 8 non-addressings; no open opposition. In our first draft we had written some models as opposing. But that was the "counter-thesis" section our question asked of every answer: models had written objections to their own theses, not opposed the thesis. So we had the real objection written by a separate counter-voice.
• Independence is less than we thought. Of eight answers, four came from the same model family (Claude) and three from another (Gemini). Asking similar models the same question doesn't yield as many independent views as there are voices. An unverified sentence had also entered our question; we removed it from the piece, but it may have affected the answers.
• We changed one recommendation. Two perception questions and the cohort that has no account yet were added to the baseline measurement (second recommendation). Both views will be tested most there.

What we recommend

Five recommendations, each with its first-year measure. The common point: most rules are already written; what's missing is data showing whether they work.

1. Make the token specification open and auditable

   The token should be single-use, short-lived, unlinkable across platforms and verifiable offline. Fix the form at two bits (15+, 18+). Publish the specification and sample code.

   First-year measureWas the guide in Article 12 published before 1 November? Did an independent audit answer "no" to "can two tokens on two platforms be linked to the same person"?

2. Take the baseline before entry into force; measure exposure, not accounts

   An independent panel (a university or the Türkiye Statistical Institute (TÜİK)) should measure before 1 November and repeat at months 6 and 12: use at ages 10–14 and 15–17, night-time use, use of out-of-scope apps, the "my family doesn't know" share and sleep.

   Add two perception questions: "Do most of your peers use social media?" and "Does someone without an account miss out?" Track separately the 10–12 cohort that has no account yet; the age limit's effect shows there first. Without a baseline no claim of success or failure can be measured.

   First-year measureWas the baseline published? How much did use fall at 12–14? (Comparison: about 4–5 points in Australia.) Did the age of first account change?

3. Make Article 8 testable from outside

   The ban on infinite feeds, the ban on personalisation, private accounts and the live-streaming restriction are already written. Now comes enforcement: keep a scorecard by platform and article with "mystery shopper" test accounts, published every three months.

   First-year measureArticle compliance rate on the five largest platforms. Complaints of unwanted contact from strangers to accounts of ages 15–17.

4. Make false rejection and exclusion visible

   Open a second path outside e-Devlet (the school administration or PTT, the national postal service). Publish monthly the number of those who can't get a token and those who appeal; the annual statistics in Article 10 are not enough for this.

   First-year measureToken failure rate. Median time to resolve appeals.

5. Measure handover and migration honestly; bring in the school

   Manage token sharing not by prohibition but by tracking its rate. Give parents and students a short, tested programme at school: how to open the parental tools in Article 9, what changes in the child's account.

   First-year measureIn the survey, the questions "whose token did you enter with" and "which app did you move to." Number of schools joining the programme.

What would change our mind?

A thesis's worth lies in saying up front what would refute it. For each of the six theses this measure is written on its own card: the highlighted boxes in the thesis cards. The same holds for views outside the cards; if the data below arrives, we withdraw or soften them.

• Ifin Australia's month-12 data the perception "most of my peers use it" has fallen markedly and account opening has fallen in the newly starting 10–12 cohort,Thenwe underrated the age limit's effect.
• Ifin Türkiye, with high compliance with Article 8 and no change in the under-15 account rate, harm indicators (unwanted contact from strangers, night-time use) fall,Thenit is design that protects, and the counter-voice's defence weakens.
• Ifin the first year enforcement decisions are published with their reasons and the number of children affected,Thenthe "unauditable infrastructure" concern eases.
• Ifthe baseline measurement has already been taken in public view,Thenour second recommendation isn't needed.

References

All the sources below were checked one by one. Those marked "partly" did not fully carry the claim at least one model based on that source; only the part they carry was used in the piece.

1. Official Gazette, 10 October 2026, no. 33396: regulation on social network providers and age verification. Anadolu Agency, "Regulation on age limits on social media published in the Official Gazette," 10 October 2026.aa.com.tr/tr/gundem/-sosyal-medyada-yas-sinirlamasina-iliskin-yonetmelik-resmi-gazetede/4084136
2. eSafety Commissioner (Australia), "Early days, early insights," three-month monitoring findings, 31 July 2026. The agency's press release, Medianet mirror.newshub.medianet.com.au/2026/07/early-insights-from-esafetys-comprehensive-evaluation-project/164211/
3. TechXplore, report on eSafety's three-month findings, July 2026.techxplore.com/news/2026-07-kids-social-media-esafety.html
4. Stanford FSI, "Australia's Age Ban, Three Months In."fsi.stanford.edu/news/australias-age-ban-three-months-what-were-learning
5. Australian government, Online Safety Amendment (Social Media Minimum Age) Act 2024 and age assurance trial fact sheet, July 2025.infrastructure.gov.au/sites/default/files/documents/social-media-minimum-age-and-age-assurance-trial-fact-sheet-july-2025.docx
6. French National Assembly, written question on Law No. 2023-566 on "digital majority."partlyquestions.assemblee-nationale.fr/dyn/17/questions/QANR5L17QE5149.pdf
7. US 11th Circuit Court of Appeals, CCIA & NetChoice v. Uthmeier (Florida HB 3), ruling of 25 November 2025.thecapitolist.com/wp-content/uploads/2025/12/11th-Circuit_CCIA-NetChoice-v.-Uthmeier_HB-3_Ruling_Nov-25-2025.pdf
8. Spectrum News, permanent court block of Ohio's social media law, 17 April 2025.spectrumnews1.com/oh/columbus/news/2025/04/17/ohio-social-media-law-permanently-struck-down-by-judge
9. The Register, start of age checks under the UK Online Safety Act, 28 July 2025.partlytheregister.com/2025/07/28/uk_vpn_demand_soars/
10. European Commission, guidelines on the protection of minors under the Digital Services Act, 14 July 2025.partlydigital-strategy.ec.europa.eu/en/library/commission-publishes-guidelines-protection-minors
11. European Commission, age-verification solution blueprint, 14 July 2025.partlydigital-strategy.ec.europa.eu/en/factpages/blueprint-age-verification-solution-help-protect-minors-online
12. Choi, Cho, Lee, Kim and Park (2018), on South Korea's shutdown policy, Journal of Adolescent Health 62(5): 548–555.ir.ymlib.yonsei.ac.kr/handle/22282913/162341
13. The Korea Herald, "Shutdown law shuttered," editorial, 16 November 2021. The National Assembly passed the amendment to the Youth Protection Act, repealing the night-time gaming ban.koreaherald.com/article/2725997
14. Sixth Tone, China's gaming-time limit for minors, 30 August 2021.sixthtone.com/news/1008398/china-limits-minors-to-gaming-at-most-3-hours-a-week
15. The Register, China's "minor mode" guidelines, 20 November 2024.partlytheregister.com/2024/11/20/china_minor_mode/
16. LSE Digital Futures for Children and 5Rights, "Impact of regulation on children's digital lives," 2024. LSE's announcement of the report.lse.ac.uk/news/latest-news-from-lse/e-may-2024/social-media-companies-make-significant-child-privacy-and-safety-changes-as-a-result-of-legislation-new-report
17. The Record, ID-photo leak at the contractor reviewing Discord's age appeals, 9 October 2025.therecord.media/discord-government-docs-exposed-breach
18. UK government, government response to the "Growing up in the online world" national consultation, July 2026. Decision to ban the service for under-16s; entry into force expected in spring 2027.gov.uk/government/consultations/growing-up-in-the-online-world-a-national-consultation/outcome/growing-up-in-the-online-world-government-response-july-2026
19. European Parliament, "Children should be at least 16 to access social media, say MEPs," press release, 26 November 2025. Non-legislative report; 483 yes, 92 no, 86 abstentions.europarl.europa.eu/news/en/press-room/20251120IPR31496/children-should-be-at-least-16-to-access-social-media-say-meps
20. Ofcom, "Keep underage children off your platforms, Ofcom tells tech firms," 12 March 2026.ofcom.org.uk/online-safety/protecting-children/keep-underage-children-off-your-platforms-ofcom-tells-tech-firms
21. Béjar and Cybersecurity for Democracy, Fairplay, Molly Rose Foundation, ParentsSOS, "Teen Accounts, Broken Promises: How Instagram is failing to protect minors," September 2025. Report by child-safety advocacy organisations.fairplayforkids.org/wp-content/uploads/2025/09/Teen-Accounts-Broken-Promises-How-Instagram-is-failing-to-protect-minors.pdf
22. Carpenter, C. and Dobkin, C. (2011), "The Minimum Legal Drinking Age and Public Health," Journal of Economic Perspectives 25(2): 133–156.aeaweb.org/articles?id=10.1257/jep.25.2.133
23. eSafety Commissioner (Australia), "Early days, early insights," full evaluation report, August 2026, Figures 8 and 9.esafety.gov.au/sites/default/files/2026-08/Early-days-Early-Insights-SMMA-Evaluation-Understanding-experiences-of-social-media-age-restrictions-August2026.pdf
24. US Federal Trade Commission (FTC), staff report on the data practices of large social media and video platforms, press release, 19 September 2024.ftc.gov/news-events/news/press-releases/2024/09/ftc-staff-report-finds-large-social-media-video-streaming-companies-have-engaged-vast-surveillance
25. Abrahamsson, S. (2026), "Smartphone Bans, Student Outcomes and Mental Health," Journal of Human Resources.partlydoi.org/10.3368/jhr.0224-13403R2
26. Orben, A., Przybylski, A. K., Blakemore, S.-J. and Kievit, R. A. (2022), "Windows of developmental sensitivity to social media," Nature Communications 13: 1649.nature.com/articles/s41467-022-29296-3
27. Bursztyn, L., Handel, B. R., Jiménez-Durán, R. and Roth, C. (2025), "When Product Markets Become Collective Traps: The Case of Social Media," American Economic Review 115(12): 4105–4136.aeaweb.org/articles?id=10.1257/aer.20231468
28. Haidt, J. and Iyer, R., "Why Every Country Should Set 16 (or Higher) as the Minimum Age for Social Media Accounts," After Babel, 13 January 2026. Advocacy essay.afterbabel.com/p/why-every-country-should-set-16
29. Friedman, A. S., Buckell, J. and Sindelar, J. L. (2019), "Tobacco-21 laws and young adult smoking: quasi-experimental evidence," Addiction.pubmed.ncbi.nlm.nih.gov/31342591/

How this article was made

Working on a case? Send it to the studio →

How this piece was made

The topic was handled with a Spark round: the same question was asked separately of AI models that couldn't see one another's answers, and 8 independent answers came back. Then a separate model was given the task of writing the strongest objection to the view the answers agreed on. The sources in all answers were checked one by one. Each model's stance on each thesis was also measured by quotation from its answer; the agreements and divergences were combined in this piece.

Date

October 10, 2026

Byline

Atlas Studio

Synthesis and writing

• Anu supervising mind; usual model Claude Fable 5.1. Finished on Claude Opus 5.5 at the end of this round because of quota

Counter-voice

• Claude Opus 5.5 Anu, red team

Independent voices

• Antigravity agent: Gemini 3.1 Pro · Gemini 3.8 Flash
• Grok Bot agent, Mac bridge: Grok (version not measured)
• Gemini app 3.1 Pro, extended thinking
• Claude Opus 5.5 headless answer
• Studio agents: Enki · Enlil · Kalfa two on Claude Opus 5.5 mounts, one Claude on the work computer

Final check and publication decision

Human editor

8independent answers

1counter-voice

48model–thesis pairings measured

0open objections

86citations checked

78citations found

53fully carry the claim

24partly carry it

9citations not used (8 not found, 1 did not carry the claim)

No fabricated source turned up; a few citation errors were corrected. Citations that couldn't be found or didn't carry the claim didn't enter this piece.

← All posts