Background for the World's First Social Media Ban for Children Was Edited by ChatGPT. The Company Denied It Until Metadata Was Found

Australia became the first country in the world to ban social media for children under 16. In an expert study worth 3.48 million Australian dollars that supported the law, The Guardian's analysis found six references to academic works with errors, including ones that do not exist. The firm that conducted the study initially denied using artificial intelligence and admitted to it only after journalists found ChatGPT metadata in four of the references. The Guardian ↗

Six References That Don't Add Up

Most of the study is based on its own tests of existing technologies. The problem is in one chapter on new technologies that cites expert articles about possible new ways to verify age. It is precisely in this chapter that The Guardian's analysis found six faulty references. The Guardian ↗

The errors take several forms. DOI identifiers that point to works that do not exist. DOIs that point to different articles than those cited. And combinations of author names, journal names, and publication years that don't match any known source.

The most specific case is dated. The study references a work that its authors accessed in March 2025, even though it was not publicly available until June. Moreover, one of the actual authors of that work confirmed to the newspaper that the person the study lists as its main author never worked on it and that the study incorrectly summarizes the work's content.

A Correction That Had Its Own Errors

When journalists asked about the errors, a spokesman for Age Check Certification Scheme denied using artificial intelligence. "We did not use AI in creating the report or the cited materials," he stated, adding that every reference was verified as genuine and relevant.

The firm then provided clarifications and new references. These also contained errors, because the years, author names, and journal titles in the revised list did not match the original citations. In one reference that could not be found in the cited journal, the firm cited a study from Monash University. In the document it sent as proof, the cited author was not present.

The firm admitted to using AI only when the newspaper found metadata in four references in two parts of the report identifying ChatGPT as the source. It argues that this was how it disclosed its use of the model, and since the references were correct according to it, it sees no problem. The model, it says, was used to rewrite some paragraphs more concisely.

A Handful of Errors, or a Problem?

The Communications Department stated it was addressing the concerns. At a Senate hearing on August 14, its officials said they had met with the firm and it explained the errors by saying that references that worked at the time of publication had since broken. The officials themselves did not verify this, saying that a retrospective check would be very demanding, and spoke of a handful of errors on 26 pages of citations in a thousand-page report. The Guardian ↗

Independent Senator Fatima Payman sees it differently and points to last year's case. According to her, the government should demand an apology and refund from the provider. The Guardian ↗

That precedent is concrete. Last October, Deloitte returned part of its fee to the Australian department for a contract worth approximately 440,000 Australian dollars. Its report on automated sanctions in social benefits contained references to non-existent works generated by the GPT-4o model, and among the fabricated sources were studies attributed to an actual law professor who never wrote them. Fortune ↗

Why this particular chapter is risky is suggested by an experiment published last year in the JMIR Mental Health journal. The authors had the GPT-4o model write six literature reviews and found that 19.9 percent of all citations were completely fabricated. Of those that appeared genuine at first glance, another 45.4 percent had bibliographic errors, most often invalid or incorrect DOI identifiers. Reliability decreased for topics with less research coverage. This is psychiatry, not age verification, so it is not evidence for the Australian case. However, the type of error and the condition under which it arises is the same: invalid DOI in a topic with limited literature. EurekAlert (JMIR Mental Health) ↗

Why This Is Not Just an Australian Problem

• 21.08.Finland shows what a phone ban looks like after a year of operation — including a side effect. Helsingin Sanomat reports that new "phone pouches" have dramatically changed school breaks, and in the same headline adds the other half: some students remain alone during breaks. Sweden's ban from the early wave can thus be compared to a country that has already gone through it. Helsingin Sanomat ↗
• 21.08.And from the American lawsuit against Meta came exactly what was missing from the Australian law — a sworn witness. Estonian Postimees reports that while Meta claims in court "we protect children," its former engineer testified from the witness stand that safety tools were actually designed to fail. This is so far testimony from one side of the dispute, not a court conclusion. Postimees ↗

• 21.08.Nordic countries are preparing a joint policy on social media for children. Icelandic public broadcaster RÚV reports that a Pan-Nordic line is being prepared on how children and young people use social media — precisely the type of regulation for which Australia has just provided a disputed expert basis. RÚV ↗
• 21.08.Sweden, meanwhile, is not beginning regulation but launching it. The nationwide mobile phone ban in schools took effect on August 1, 2026: students aged 7 to 16 hand over their phones by the end of class, with exceptions only for health reasons, and the state allocated 95 million kronor for this year and 100 million for next year. A Dagens Nyheter commentator summed it up on the first day back to school with a sentence that Swedish children are finally forced to be without their phones. Dagens Nyheter ↗

Europe is not following the Australian law from a distance but is preparing its own version. European Commission President Ursula von der Leyen admitted in May 2026 to a "delay of social media" for children and announced a possible legislative proposal. Euronews ↗

The Commission's expert panel submitted its recommendations on July 13, 2026, and proposed an age-differentiated approach. Children under 13 should have access to social media only with parental consent or for educational purposes, according to it. Panel Report (European Commission) ↗ The Commission is also pressing member states to make the European age verification application available by the end of 2026. European Commission ↗

As of May 11, 2026, 23 of 27 member states were preparing or considering their own regulations, but they differ in age limits and which services should be covered. interface ↗

How difficult it is to write such a law was shown by France. A ban on social media for those under 15 was to take effect in September, but the Constitutional Council struck it down on August 14 as an infringement of freedom of expression that is neither proportionate nor necessary. President Emmanuel Macron subsequently asked the Prime Minister for a legally more robust version. Conseil constitutionnel ↗

The Australian test, however, is not a marginal document for Europe. It was the first major attempt to measure whether age estimation technologies work at all, and Brussels directly refers to the Australian model. Biometric Update ↗

None of this says that age verification does not work or that the Australian law is bad. The errors are in one chapter of a thousand-page report and the firm insists the model did not fabricate the citations. However, Christian Downie, an expert on regulation from Australian National University, points out what matters. If false citations appear in government documents, whether from AI or not, it can lead to bad decisions and undermine trust in institutions. And without it, a law restricting children is much harder to justify. The Guardian ↗