According to The Guardian, the authors of a report examining the technology behind Australia’s social media ban have admitted using ChatGPT during the editing process, but denied that multiple citation errors in the report were caused by artificial intelligence “hallucinations.”
The A$3.48 million test of age-verification technologies (approximately RMB 16.629 million at current exchange rates) was conducted last year by the UK-based Age Check Certification Scheme (ACCS). It primarily tested various technologies that social media platforms could potentially use to enforce Australia’s ban on social media access for users under 16.
Australian Communications Minister Anika Wells previously praised the report, saying it demonstrated “many effective options” and helped pave the way for the social media ban to formally take effect last December.
Most of the report was based on the authors’ practical testing of various age-verification technologies currently available. However, one chapter was dedicated to “emerging technologies” and cited several academic papers describing new methods that could potentially be used for age verification in the future.
The Australian Senate is currently conducting an inquiry aimed at further strengthening laws related to the social media ban. During the inquiry, a submission identified at least two citation errors in the report. It said the “citations appear to have been generated by AI hallucinations rather than being based on genuine sources.”
When Guardian Australia asked ACCS about the matter this month, an organization spokesperson initially denied that AI had been used in the report, but later acknowledged that the team had used AI to rewrite some passages to make them more concise.
The spokesperson said: “We did not use AI to generate the report or any of the cited material … Every item was checked to confirm that the links and reports genuinely existed and were relevant to the specific issue for which they were cited.”
However, Guardian Australia’s own analysis of the report found six erroneous references in that chapter alone.
The errors included digital object identifiers (DOIs—a unique string of letters and numbers corresponding to a particular academic paper) linking to papers that did not exist; DOIs pointing to the wrong papers; combinations of author names, journal titles and publication years that did not match any known literature; and DOI links to papers that did not support the conclusions claimed in the report.
The spokesperson subsequently provided further clarification about the research on which ACCS had relied and supplied new references, but that response also contained errors.
In a list of academic papers supplied by ACCS to correct the report’s original citations, the publication years, authors and journal names of some papers did not match the citations originally listed in the report.
One paper cited by ACCS was reportedly accessed in March 2025. However, one of the study’s lead authors confirmed to The Guardian that the paper was not publicly available until it was formally published in June 2025. He also confirmed that “Jamil,” whom ACCS had originally listed as the paper’s lead author, had never been involved in the paper, and that the ACCS report had incorrectly summarized its contents.
Regarding another reference, “Weber et al. 2011,” which could not be found in the cited journal, the spokesperson further explained that the work had been cited by a Monash University study and provided that study to The Guardian. However, the Monash University paper supplied by ACCS contained no reference attributed to anyone named Weber.
The spokesperson did not acknowledge that AI had indeed been used in the report until The Guardian discovered that four links in Sections E and K contained metadata showing that they had originated from ChatGPT.
The spokesperson maintained that because the links retained ChatGPT metadata, the organization had effectively disclosed its use of AI. The spokesperson also said that the links themselves were correct and therefore ACCS did not consider this to be an issue.
He said AI had been used only to make the wording of some passages more concise, and had not been used to conduct original research or generate the report.
“All links were checked and verified, and if there are any inadvertent errors, I apologize, but the entire process was verified by humans.”
The Australian federal government has previously taken a negative view of contractors using AI to prepare reports for the government. Deloitte refunded part of a A$440,000 government contract (approximately RMB 2.103 million at current exchange rates) last year after errors were found in its report, and the consulting firm admitted that it had used AI.
Questions submitted by The Guardian to Wells’ office were subsequently referred to her department. A department spokesperson said last Friday: “The department is reviewing the matters outlined below and will engage with ACCS as appropriate.”
At a Senate inquiry hearing held last Friday local time, departmental officials said they had met with ACCS. ACCS claimed that the errors in the report arose because some links that had originally worked later became inaccessible. However, officials had not independently verified this claim and said it would be “very difficult” to go back and check the links.
Officials said that in a 1,000-page report, only 26 pages contained citations and there were merely “a small number of errors.”
Sarah Vandenbroek, the department’s first assistant secretary, said: “ACCS has assured us that all links were checked when the document was published and were accessible at that time. It is therefore possible that something changed in the interim. They have also assured us that even if some links are now inactive, the underlying original documents still exist and the citations remain valid.”
Christian Downie, a professor at the Australian National University’s School of Regulation and Global Governance, has previously studied submissions received by the government that contained AI hallucinations. He said that if reports or submissions received by the government were found to contain fabricated citations, regardless of whether the errors were caused by AI, they could lead the government to make poor decisions and “undermine public confidence and trust in government institutions.”
He said: “People should expect the government to be considering this issue. In the future, contractual provisions and corresponding penalties may be needed to deter this type of conduct.”
Independent Senator Fatima Payman said: “The controversy over Deloitte’s AI-slop-filled report last year should have put an end to government contractors casually citing sources in reports.” She added: “The government must demand that the contractor responsible for this report apologize and issue a refund.”