A key document on age verification technology testing, used by Australian authorities to justify a social media ban for teenagers, contains references to non-existent scientific publications. This conclusion was reached by journalists from The Guardian, who examined the report’s text.
The tests were conducted by the British organization Age Check Certification Scheme (ACCS), which spent 3.48 million Australian dollars (~$2.5 million) on them. Experts evaluated over 60 solutions from 48 providers, and in August 2025, they submitted a final document to the government, spanning about a thousand pages across ten volumes.
At the time, Communications Minister Anika Wells stated that the data confirmed the existence of “numerous effective tools” for age verification on platforms. The ban for users under 16 came into effect on December 10, 2025.
Six References with Errors
The main part of the report focuses on practical tests of existing solutions. Issues were found in the chapter on prospective developments, where the authors refer to scientific literature.
The discrepancies were first pointed out by the author of a submission to a Senate committee currently reviewing amendments to tighten the social media ban for users under 16. The document claims that at least two references resemble typical AI “hallucinations” rather than citations of real sources.
Guardian journalists verified the data and found six violations in the same chapter, grouped into four categories:
- fictitious DOI;
- DOIs redirecting to unrelated works;
- author-journal-year combinations not supported by any publication;
- correct article references but lacking the claims attributed to them in the report.
For example, one article was dated by ACCS as March 2025, but it was actually published in July. In another case, a reference to Weber et al. 2011 led nowhere—it was not found in the specified journal or in the Monash University study the report authors cited.
ChatGPT in Metadata
An ACCS representative categorically denied using AI in preparing the document. According to them, the organization did not apply neural networks to the text or cited materials, and each reference was thoroughly checked for accuracy and relevance.
However, after The Guardian found four references with metadata directly indicating ChatGPT in some parts of the report, the stance had to be softened. ACCS admitted that artificial intelligence did rewrite certain sections for brevity but insisted that the research and final text were created without OpenAI’s chatbot.
Representatives of the organization also stated that ChatGPT metadata itself serves as sufficient notification of AI use, and since the links lead to the correct destinations, there should be no complaints.
However, a check of the corrected list of sources provided to the publication by ACCS revealed new discrepancies: years, authors, and journal titles differed from the original report citations.
“Handful of Errors” in a Thousand Pages
The Department of Infrastructure and Communications reported that they are reviewing the claims and intend to contact the contractor. On August 14, during Senate hearings, officials recounted ACCS’s version: according to them, the links were functional at the time of publication and “broke” later.
Sarah Vandenbroek, the department’s first assistant secretary, clarified that the department did not independently verify this version and considers such verification “labor-intensive.” However, the department emphasized that the issue concerns only a “handful of errors” on 26 pages of bibliography.
Professor Christian Downie of the Australian National University warned that false references in government documents lead to erroneous decisions and undermine trust in institutions. In his view, this situation requires clear contractual requirements and sanctions for violations.
Independent Senator Fatima Payman demanded an official apology from the government and a refund from the contractor. She reminded that a similar incident with Deloitte a year ago should have taught the authorities to pay more attention to sources. In October 2025, the consulting company had already returned part of the payments on a 440,000 Australian dollar contract, admitting the use of AI in preparing a report with errors.
In July, the Australian regulator eSafety filed a lawsuit against Telegram, accusing the messenger of refusing to remove materials promoting terrorism.
