The key document on which the Australian government based its justification for banning social media for teenagers turned out to be riddled with references to non-existent scientific works. My analysis uncovered a systemic problem: the report, prepared by the British organization Age Check Certification Scheme (ACCS) for 3.48 million Australian dollars (~$2.5 million), contains at least six gross violations in its bibliography—from fictitious DOIs to articles backdated after the fact.
The trials covered more than 60 solutions from 48 vendors, and the final document, about a thousand pages long in ten volumes, was handed to the government in August 2025. Communications Minister Anika Wells then stated that the data confirmed the existence of "many effective tools" for age verification. The ban itself for users under 16 took effect on December 10, 2025—and now its legitimacy hangs by a thread.
Six references that unravel the argument
The problems are concentrated in the chapter on promising developments, where the authors turn to scientific literature. The first to point out the inconsistencies was the author of a submission to the Senate committee reviewing amendments to tighten the ban. My check identified four categories of errors: fictitious DOIs, references to unrelated works, non-existent "author-journal-year" combinations, and correct references but with distorted claims.
For example, one article is dated March 2025, although it was actually published in July. Another reference to Weber et al. 2011 leads nowhere—it appears neither in the cited journal nor in the Monash University study. These are not random typos but a systemic trace of generative models.
ChatGPT in the metadata: denial and admission
Representatives of ACCS initially categorically denied the use of AI, claiming that every reference was manually verified. However, after direct indications of ChatGPT were found in the metadata of four references, the position had to be softened. The organization admitted that the neural network rewrote individual fragments for brevity but insisted that the research and final text were created without the chatbot's involvement.
Moreover, ACCS stated that the ChatGPT metadata itself constitutes sufficient notice of AI use. This is a dangerous precedent: if a marker in a file counts as disclosure, then verification of accuracy becomes a formality. The corrected list of sources, provided to the publication, revealed new discrepancies in years, authors, and journal titles.
A "handful of errors" or a crisis of trust?
The Department of Infrastructure and Communications said they are reviewing the claims, while officials at Senate hearings relayed ACCS's version: the references allegedly "broke" after publication. First Assistant Secretary Sarah Vandenbroek admitted that the department had not verified this version, considering the check "labor-intensive." The department insists that this is a "handful of errors" across 26 pages of bibliography.
Professor Christian Downie from the Australian National University warns: false references lead to flawed decisions and erode trust in institutions. Independent Senator Fatima Payman demands an apology and a refund, recalling the precedent with Deloitte, which in October 2025 returned part of its payment for a report with AI errors.
My verdict: this case is not just a technical glitch but a symptom of an era when the speed of content generation outpaces responsibility for its accuracy. If governments make regulatory decisions based on "hallucinating" documents, we risk getting laws built on sand. Contracts with contractors must include strict sanctions for using AI without verification; otherwise, trust in institutions will erode with every such report.