A report that helped justify Australia’s under-16 social media ban contains academic citations that link to papers which do not exist, according to a Guardian Australia investigation. The authors conceded ChatGPT was used in editing the document — but denied the citation errors were caused by AI.
The $3.48 million age-assurance technology trial, run by the UK-based Age Check Certification Scheme (ACCS), tested technologies that social media platforms could use to verify users’ ages. Communications minister Anika Wells heralded the report as showing “many effective options” for age checking, and it paved the way for the ban to take effect in December 2025.
Now a Senate inquiry has heard that at least six references in the report’s “emerging technologies” chapter contain errors: DOIs linking to non-existent papers, author-journal-year combinations that match no known publication, and citations that misrepresent what the original papers actually said.
What the Investigation Found
The Guardian’s analysis identified six references with fabricated metadata in a single section of the 1,000-page report. The errors include Digital Object Identifiers — unique strings that point to specific academic articles — that resolve to nothing, or to entirely different papers than the ones cited.
When first asked about the allegations, an ACCS spokesperson denied using AI. After the Guardian identified four links containing metadata that traced back to ChatGPT, the spokesperson conceded AI had been used — but only to “rewrite paragraphs more succinctly.” They maintained the links were correct and had been verified by humans.
That position did not survive scrutiny. One article ACCS claimed was accessed in March 2025 was not publicly available until June 2025, according to the paper’s actual lead author. The person ACCS listed as lead author — “Jamil” — had never been on the paper. The report’s summary of the paper’s findings was, in the author’s words, incorrect.
Another reference, “Weber et al. 2011,” could not be found in the journal cited. ACCS said it came from a Monash University study they provided to the Guardian. There was no reference attributed to Weber in that document either.
A Pattern, Not an Isolated Case
This is not the first time an Australian government contractor has been caught using AI to produce official reports. Deloitte refunded part of a $440,000 contract last year after errors were identified in a report the firm admitted was produced with AI assistance.
The pattern extends well beyond Australia. South Africa’s government withdrew an AI policy paper after it was found to contain hallucinated citations. USENIX rejected 21 academic papers for the same problem. Even EY pulled a study on AI in consulting after discovering fabricated references.
What connects these cases is the same failure mode: a tool that generates plausible-sounding text, including citations that look real but are not. The problem is not that AI was used — it’s that the output was treated as finished work without verification.
”A Handful of Errors”
At Friday’s Senate hearing, department officials said ACCS claimed the broken links had previously worked and that something “may have changed in the interim.” The officials have not independently verified this, stating it would be “very difficult” to check.
Sarah Vandenbroek, the department’s first assistant secretary, characterised it as “a handful of errors” in 26 pages of citations across a 1,000-page report. ACCS assured officials the links were all checked at the time of publication.
Prof Christian Downie, from the Australian National University’s school of regulation and global governance, told the Guardian that false citations in government reports — AI or otherwise — could lead to bad decisions and “erode public confidence and trust” in institutions. He suggested contractual penalties may be needed to discourage the practice.
Independent senator Fatima Payman was blunter: “The saga of last year’s Deloitte AI slop report should have marked the end of slack referencing in reports by government contractors.”
The Bigger Problem
The age-assurance report was not a routine document. It was the technical foundation for a law that prevents anyone under 16 from using social media in Australia — a law other countries, including New Zealand, are watching closely as a potential model.
If the emerging-technologies chapter contained fabricated citations, how confident can policymakers be in the rest of the report? The ACCS’s initial denial, followed by a partial admission only after being confronted with technical evidence, does not inspire confidence in their verification processes. The claim that humans checked every link sits awkwardly alongside the discovery of ChatGPT metadata embedded in the document itself.
This connects to a broader question we’ve explored before: when AI-generated content enters the regulatory pipeline, who is responsible for the consequences? Australia’s social media ban affects millions of young people. The technical report supporting it should be held to a higher standard than a ChatGPT editing pass.
For New Zealand, which has been comparing its own approach to Australia’s, the lesson is straightforward. If you’re going to cite expert research to justify restricting access to platforms, the research needs to actually exist.
