UK contractor behind Australia's teen social media ban report admits AI use

Australia's under-16 social media ban, in force since December 2025, was justified in part by a A$3.48m government-commissioned report that Guardian Australia's reporting shows contains at least six broken citations in its "emerging technologies" chapter: links to papers that do not exist or do not say what the report claims they say.

The Age Assurance Technology Trial (AATT) was commissioned by the Australian government and run by the UK-based Age Check Certification Scheme (ACCS). Communications Minister Anika Wells publicly framed it as showing "many effective options" for age checking before the ban took effect, and the report was the principal published evidence base for the policy.

The dispute does not concern the bulk of the report, which documents the age-verification testing the government relied on. It concerns Part E, a chapter on emerging technologies such as biometric age estimation, device-based inference, and behavioural signals. Age assurance is the umbrella term for the systems platforms use to estimate or verify that a user meets a minimum age, ranging from ID uploads and selfie-based biometric checks to passive inference from usage patterns. Part E surveys the methods that go beyond ID upload.

According to Guardian Australia's analysis, the Part E chapter contains at least six reference errors: digital object identifiers that resolve to nothing, DOIs that point to the wrong paper, author-year-journal combinations that match no known academic reference, and studies that do not contain the claims the report attributes to them. A DOI is a permanent identifier assigned to a published academic paper; if the identifier does not resolve, the paper either does not exist or is not the paper the citation claims.

A submission to the Australian Senate's inquiry into the age-assurance regime alleges that the citations "appear to be AI hallucinated, rather than being based on real sources."

When Guardian Australia first reported the citation problems, ACCS denied that AI had been used in the report. The organisation later conceded, after the Senate submission was lodged, that AI had been used to "rewrite paragraphs more succinctly" in Part E. ACCS continues to deny that AI generated the citations themselves and says each citation was checked as a genuine link.

ACCS also published a follow-up document of "clarifications and new references" in response to the criticism. That document contained its own errors: years, authors, and journal titles in the replacement references do not match the originals, according to Guardian Australia's analysis.

The verification gap is the operative failure mode. AI-assisted drafting and AI-fabricated citation are not the same act, and the report's response arc conflates them. If AI was used only to compress prose, the underlying citations should have been independently verifiable before publication. If the citations cannot be independently verified, the rewriting defence does not address how they entered a government-commissioned report in the first place.

The AATT report is publicly hosted on the Department of Infrastructure's publication hub. Part A covers the trial methodology; Part E is the chapter under dispute.

A Guardian Australia follow-up piece in August 2025 had already noted that "some errors" in the report were "inevitable" given the trial's compressed timeline, and ACCS's own preliminary findings release in June 2025 flagged the trial's interim status.

The Department of Infrastructure has not, as of Guardian Australia's reporting, indicated whether the report is being formally withdrawn, corrected, or re-issued. Minister Wells' office has not commented on the Senate submission. The Senate inquiry is scheduled to continue hearing submissions on the dispute.

The constructive question the inquiry now faces is not whether AI was used; that is conceded. It is what verification standard should apply to AI-assisted commissioned research that informs a children's-safety policy affecting millions of accounts. The current ACCS position, that AI rewrote prose but did not generate the citations it could not, on correction, reproduce, does not, on the available record, answer that question.