TL;DR
ACCS — the Age Check Certification Scheme, a UK-based body — ran the $3.48m technology trial that underpinned Australia’s under-16 social media ban. Its report contains citation errors, including links to papers that do not exist. The organisation first denied using AI, then admitted ChatGPT had been used for editing after journalists found its metadata embedded in four of the report’s links.
What was wrong with the references
The Guardian found six flawed references in a single section covering emerging age-check methods. The problems ranged widely: digital object identifiers resolving to nothing, identifiers pointing at the wrong paper, and author-journal-year combinations matching no known publication. Some cited papers existed but did not support the claim attached to them.
When ACCS supplied corrections, those contained errors too — years, authors and journal names that failed to match the originals. One paper it claimed to have consulted in March 2025 was not publicly available until June, and a lead author confirmed the person ACCS named had never worked on it, and that the summary of their research was wrong.
The organisation maintains AI was used only to tighten prose, never to originate research or generate citations, and argues that leaving ChatGPT metadata in the links amounted to disclosure. Department officials told a Senate inquiry that ACCS attributed the failures to links breaking after publication. They have not verified that, calling it very difficult to check retrospectively.
Why UK firms should read this closely
The certifier is British, and the UK’s own age assurance regime under the Online Safety Act rests on exactly this class of independent certification body. A professional services organisation whose product is verification has been caught unable to verify its own footnotes.
Australia’s government has form here: Deloitte refunded part of a $440,000 contract last year after admitting AI use in a report containing errors. Professor Christian Downie of the Australian National University argues false citations in evidence to government risk bad decisions and eroded institutional trust, and expects contracts to start carrying penalties.
Looking forward
The defence on offer — the tool only rewrote paragraphs — sidesteps the actual failure. Editing assistance does not fabricate a digital object identifier. Any UK firm submitting evidence to a regulator or consultation should assume that citations are now the first thing a hostile reader checks, because they are the cheapest place to catch an unverified claim.