Accscheme.com
A government-commissioned report produced to demonstrate the technical feasibility of Australia’s world-first under-16 social media ban was partly written using ChatGPT, contains at least six fabricated academic citations, and the organization that produced it initially lied about using AI at all. A Senate committee heard evidence of the fraud today, raising sharp questions about whether the legislation — now being copied by more than 25 countries — rested on manufactured scholarly proof.
The report in question was produced by the Age Check Certification Scheme (ACCS), a UK-based nonprofit that was contracted to lead Australia’s A$3.48 million trial contract — the testing phase the government cited as evidence that platforms could reliably block minors from social media. The trial carried a contract value of A$3.48 million (approximately $2.47 million USD).
The finding matters beyond the immediate scandal. Australia’s Senate is currently deliberating an enforcement bill now before Senate that would nearly double the maximum penalty for platform non-compliance — from A$49.5 million (approximately $35.1 million USD) to A$99 million (approximately $70.3 million USD). The committee’s report is due August 25. Those deliberations now proceed with the knowledge that the government’s own technical case for the ban’s feasibility was contaminated.
How Did Fabricated Citations End Up in a Government Report?
Large language models like ChatGPT do not retrieve facts. They generate statistically plausible text based on patterns in their training data — and they have no internal mechanism to distinguish a real paper they encountered during training from a plausible-sounding one they are inventing from scratch. When asked to find supporting research on a topic, they produce what looks like a citation: an author name, a journal title, a publication year, and a Digital Object Identifier (DOI) — the unique handle used to locate and link to academic papers. The problem is that any of those components can be entirely fabricated, and the DOI may point to nothing, to a paper that does not exist, or to a real paper that says something entirely different from what the citation claims.
A January 2026 analysis by GPTZero of papers accepted to NeurIPS 2025 — the premier AI research conference, with three to five expert peer reviewers per paper — found 100 confirmed hallucinated citations across 53 accepted papers. If expert peer review misses AI-fabricated citations at that frequency in a room full of professional AI researchers, procurement officers reviewing a government-contracted age assurance report have near-zero chance of catching them without dedicated verification tools.
Academic research published in 2025 found that citation hallucination rates climb sharply — to between 28 and 55 percent — on niche or recent topics, compared to roughly 6 percent for well-documented subjects. The “emerging technologies” chapter of an age assurance trial report written in 2024-2025 sits squarely in that high-risk zone: it covers a narrow and rapidly evolving field with a limited body of peer-reviewed literature.
Guardian Australia’s citation analysis identified six flawed references in that chapter. Some DOIs were linked to papers that do not exist anywhere in academic databases, complete with fabricated authors, journals, and publication years. At least one DOI pointed to a real paper — but not one that said what the ACCS report attributed to it.
The Cover-Up: Denial, Admission, and a New Set of Errors
When Guardian Australia first approached ACCS about the citation problems, a spokesperson initially denied AI use entirely — stating that no artificial intelligence had been used to produce the report.
That denial collapsed under scrutiny. ACCS subsequently admitted that ChatGPT had been used to rewrite paragraphs, and that metadata from the AI tool was discoverable in four links across two sections of the document. The organization argued that the presence of embedded metadata constituted adequate disclosure — a claim that drew immediate skepticism from senators given that ACCS had, seconds earlier, denied AI use entirely.
The manual-verification claim that ACCS fell back on was quickly undermined. When the organization attempted to correct the flagged citations after Guardian Australia’s inquiry, it produced fresh correction errors. Among them: ACCS claimed one paper had been accessed in March 2025, months before its actual June 2025 publication date — a fact confirmed by the paper’s own lead author.
The ministerial department told today’s Senate inquiry that it had met with ACCS following Guardian Australia’s investigation. According to the department’s briefing, ACCS explained the presence of faulty links — but did not, the department said, disclose that the problem was fabrication.
A Pattern Australia Has Already Seen
Independent Senator Fatima Payman drew an explicit parallel to a domestic precedent at today’s hearing. In October 2025, consulting firm Deloitte partially refunded A$440,000 government contract after an academic discovered that a report it had produced for Australia’s Department of Employment and Workplace Relations was riddled with AI-generated errors — including citations to nonexistent studies and a fabricated quote attributed to a federal court judgment.
Deloitte eventually acknowledged using Azure OpenAI’s GPT-4o in drafting portions of the welfare compliance report. The corrected report removed fabricated references, published September 26, 2025, while retaining the same policy recommendations. Deloitte refunded approximately A$97,000 (approximately $69,000 USD) — roughly the final installment of its fee — and maintained that the substance of the report’s conclusions remained sound.
The Australian Greens flagged at the time that Deloitte had engaged in conduct that a first-year university student would be in deep trouble for. A broader pattern has since emerged. GPTZero’s investigations found that PwC Middle East produced four government and public-sector reports between 2024 and 2026 with hallucinated citations, including one promoting a framework called “Citizen Pulse” for which little public evidence of existence could be found. Deloitte’s Canadian arm produced a A$1.6 million (approximately $1.14 million USD) health workforce plan for Newfoundland and Labrador that cited researchers on nonexistent papers.
The thread connecting every case is the same: AI tools used to bulk-produce the appearance of scholarly rigor, without anyone verifying that the underlying sources existed.
Can Senate Deliberations Trust the Evidence Base?
Christian Downie, a professor at the Australian National University’s School of Regulation and Global Governance, told the Senate inquiry that fabricated citations — regardless of how they were generated — carry a corrosive effect on public policy.
False references, Downie argued, increase the likelihood of decisions being made on a flawed evidence base, and when discovered, they undermine public confidence in the policy and in the institutions that commissioned and relied on the work.
The timing compounds the problem. The ACCS trial was not peripheral to the under-16 ban — it was the government’s primary mechanism for demonstrating that age verification was technically achievable before Parliament voted on the law. Platform companies had urged Parliament to wait until the trial’s findings were available before passing the legislation; Parliament pressed ahead anyway, with the understanding that the trial would produce the technical validation in short order. The trial did produce findings. Those findings, it now appears, included scholarly support that was invented.
That structural gap — a government commissioning a single contracted report to prove the technical feasibility of proposed legislation, and then citing that report in parliamentary debate, without independently verifying its academic foundations — is the deepest problem the ACCS revelation exposes. It is not specific to ACCS. It is how evidence-based policy works when AI tools make it possible to generate an authoritative-looking citation list in minutes, and when procurement contracts contain no requirement to verify that the cited papers exist.
Communications Minister Anika Wells had praised the ACCS report, citing it as evidence of viable age verification solutions. As of the time of writing, neither the minister nor her department has issued a public statement correcting or withdrawing that endorsement. The ACCS report itself has not been retracted.
What Happened to the Ban Without the Report?
The enforcement picture that has emerged since the ban took effect in December 2025 does not depend on the ACCS report to be troubling. The eSafety Commissioner longitudinal evaluation found that 81.5 percent of Australian children under 16 were still using at least one restricted platform three months after the law took effect, compared to 85.9 percent before the ban — a reduction of roughly 4 percentage points. More than half of the 803 children surveyed said the platform they were still using had not checked their age at all.
Australia’s law requires platforms to take “reasonable steps” to prevent under-16s from holding accounts but does not specify a technical minimum — no requirement for point-of-sign-up verification, no mandate for privacy-preserving zero-knowledge proof systems, no prohibition on the document-scan approaches that concentrate sensitive identity records at breach-prone third-party vendors. That vagueness in the law’s drafting is what the age assurance trial was supposed to resolve: the trial would identify which technologies worked and how. If the chapter of the trial report dedicated to “emerging technologies” was built on fabricated scholarship, the government’s technical answer to “how do we actually verify age?” rests on thinner ground than it presented to Parliament.
The Senate’s August 25 enforcement report must now answer not only whether higher penalties will change platform behavior, but whether the law it is asking platforms to comply with ever had an independently verified technical foundation in the first place.
Is ACCS’s Conflict of Interest the Missing Context?
One element absent from the government’s framing of the ACCS trial deserves scrutiny. ACCS is not a disinterested research organization. It is a certification and standards body that tests and certifies age verification technology providers — companies with a commercial interest in the conclusion that age verification is technically feasible, achievable at scale, and ready for legislative mandate.
A trial commissioned from an industry trade body to evaluate whether the industry’s own products work, funded at A$3.48 million (approximately $2.47 million USD) with findings that would be cited as evidence in one of the most consequential technology regulation debates in Australian history, with no independent validation of the academic citations underlying its conclusions — that is a procurement and governance design problem that exists independently of whether ChatGPT was used to generate the report’s references.
The Senate inquiry has not yet indicated whether it will call ACCS to give oral evidence or seek a formal accounting from the ministerial department about the due-diligence process used when accepting and publicly citing the report.
Frequently Asked QuestionsWhat are fabricated academic citations in AI-generated reports, and why are they so dangerous in policy documents?
When large language models like ChatGPT are asked to support a claim with research, they generate fabricated but plausible citations — an author, journal, year, and DOI — without any mechanism to verify that the source exists. The result can look indistinguishable from a real citation to a non-expert reader. In a government policy report, fabricated citations allow a contractor to present invented scholarly consensus as evidence that a policy approach is technically sound, scientifically validated, or supported by research. When legislators and ministers cite that report in parliamentary debate, the invented evidence becomes part of the public record underlying a law. The danger is not that the citations are wrong in an easily checkable way — they are wrong in a way that requires dedicated verification tools and domain expertise to catch, and most government procurement processes include neither.
What specifically was wrong with the ACCS report, and has it been withdrawn?
The Guardian Australia citation analysis found six flawed references in the report’s “emerging technologies” chapter. Some DOIs led to papers that do not exist in any academic database. Some citations attributed findings to authors and journals that nowhere appear in the scholarly literature. At least one DOI pointed to a real paper that did not say what the report claimed. ACCS initially denied using AI, then admitted ChatGPT had been used to rewrite paragraphs, and when it attempted to correct the flagged citations, produced a new set of errors — including an access date for a paper that predates the paper’s actual publication. As of August 17, 2026, the report has not been withdrawn, and Minister Anika Wells has not issued a public correction of her earlier endorsement.
Why does the ACCS citation scandal matter to countries other than Australia that are now considering social media age bans?
Australia’s age assurance trial was intended to be the technical proof-of-concept that other governments would learn from — evidence that age verification is achievable at meaningful scale without unacceptable privacy costs. 25 nations following Australia’s model have enacted restrictions or tabled legislation modeled on Australia’s approach, per an Agence France-Presse tally from July 2026. If the report that demonstrated technical feasibility contained fabricated scholarly support for specific verification technologies, every country drawing on Australia’s trial findings to design its own enforcement architecture has been working from a contaminated map. The structural lesson — that a government-commissioned trial report produced by an industry trade body, with no independent citation verification, should not be treated as settled science — applies to any regulatory process that outsources its technical evidence production to a vendor with a stake in the outcome.
What practical steps can Australian legislators take now to address the evidentiary gap?
Independent verification of the ACCS report’s remaining claims is the minimum required response — specifically, a line-by-line audit of every citation in the “emerging technologies” chapter by researchers with no relationship to ACCS or the age verification industry. The Senate’s August 25 enforcement report could recommend that any future trial findings used to inform legislation be reviewed by an independent body before being cited in parliamentary debate. More broadly, Australia’s government procurement frameworks do not currently require contractors to disclose AI tool use or verify AI-generated citations — a gap the Deloitte and ACCS cases have now exposed twice within twelve months. Mandating disclosure and verification as standard contract terms would address the problem structurally rather than case by case.