A $3.48m evidence base, six bad citations, and a denial that did not survive the metadata

The authors of the technology trial underpinning Australia's under-16 social media ban have conceded ChatGPT was used in editing the report, while denying that citation errors in it were AI hallucinations ^[1]. This is not a brochure that can be quietly corrected: the communications minister, Anika Wells, presented the report as showing "many effective options" for checking ages, and it paved the way for the ban to take effect in December last year ^[4][5].

The $3.48m age assurance technology trial was run by the UK-based Age Check Certification Scheme (ACCS) ^[2], testing technologies platforms could use to comply with the ban ^[3]. Most of the report rests on the authors' own testing of available products ^[6]. The problems sit in one chapter on "emerging technologies" that cites journal papers ^[7]. A submission to an ongoing Senate inquiry into strengthening the law flagged at least two citations that "appear to be AI hallucinated, rather than being based on real sources" ^[8]. Guardian Australia's own analysis of that section found six references containing errors ^[10]: DOIs linking to papers that do not exist, DOIs pointing to the wrong papers, author, journal and year combinations matching no known reference, and DOI links to papers that did not say what the report said they said ^[11].

The sequence after that is the part worth reading twice. ACCS initially denied using AI in the report at all ^[9], stating that each citation had been "checked as genuine links and reports" and confirmed as relevant [c9b]. It then supplied clarifications and replacement references, and those contained errors too, with years, authors and journal titles that did not match the original citations ^[12]. One cited paper was recorded as accessed in March 2025, but a lead author told the Guardian it was not publicly accessible until it was published in June ^[13], roughly three months later ^[1]; the same author said the person ACCS named as lead author, "Jamil", had never been on the paper, and that the report had summarised it incorrectly ^[13]. Asked about "Weber et al. 2011", which could not be found in the journal cited, ACCS pointed to a Monash University study, which contained no reference attributed to Weber ^[14].

The concession came only after the Guardian identified four links in parts E and K of the report carrying metadata showing ChatGPT as their source ^[15]. ACCS's position is that including that metadata amounted to disclosure and, since the links were correct, there was no problem ^[16], and that checking "was all done by human verification" ^[17]. Its account of the use is narrow: rewriting some paragraphs more succinctly, not originating research ^[18].

There is a price list for this already. Deloitte refunded part of a $440,000 government contract last year after errors were identified and the firm admitted using AI ^[19]. The ACCS trial cost roughly eight times that contract ^[2].

The department said it is examining the concerns raised and will engage with ACCS as appropriate ^[20], while questions to Wells' office were redirected there ^[21]. At a Senate inquiry hearing on Friday, officials said ACCS had told them the errors were caused by links breaking that previously worked ^[22]. Two things to watch: whether the department re-verifies the emerging technologies chapter itself rather than accepting that explanation, and whether any of the $3.48m follows the Deloitte precedent back to the Commonwealth ^[19][2].