Ghost links: how AI hallucinations undermined trust in the Australian report on the social media ban
A government report that served as a cornerstone for introducing Australia's ban on social media for teenagers has been found to contain traces of systemic errors typical of generative neural networks. This concerns a key document on testing age verification technologies, prepared by the British organization Age Check Certification Scheme (ACCS). My analysis has uncovered serious discrepancies in the bibliography, calling into question the quality of the entire evidence base.
The trials cost the Australian budget 3.48 million Australian dollars (~$2.5 million). ACCS specialists tested more than 60 solutions from 48 providers, and the final document, about a thousand pages long in ten volumes, was handed over to the government in August 2025. At the time, Communications Minister Anika Wells stated that the data confirmed the existence of "many effective tools" for age verification. The ban itself for users under 16 came into effect on December 10, 2025.
Systemic Failures in the Bibliography
The main part of the report is devoted to practical trials, but problems emerged in the chapter on promising developments, where the authors refer to scientific literature. I have identified at least six violations that fall into four categories: fictitious DOIs, links leading to unrelated works, non-existent "author-journal-year" combinations, and correct references but with distorted claims.
For example, one article is dated March 2025, although it was actually published only in July. In another case, a reference to Weber et al. 2011 leads nowhere—it is found neither in the specified journal nor in the Monash University study. These are classic signs of AI "hallucinations," not the diligent work of a researcher.
ChatGPT in the Metadata
ACCS representatives initially categorically denied the use of AI. However, after traces of four references directly pointing to ChatGPT were found in the metadata, the position changed. The organization admitted that the neural network rewrote individual fragments for brevity but insisted that the research itself was created without the involvement of OpenAI's chatbot. At the same time, a check of the corrected list of sources revealed new inconsistencies: years, authors, and journal names diverge from the original citations.
The Department of Infrastructure and Transport stated that they are reviewing the claims. Officials relayed ACCS's version that the references "broke" after publication but admitted they had not independently verified this version, considering the check "labor-intensive." Professor Christian Downie of the Australian National University warns that false references in government documents lead to erroneous decisions and erode trust in institutions. Independent Senator Fatima Payman demands an apology and a refund of funds, recalling the recent incident with Deloitte, which had already returned part of its payments for using AI in a report with errors.
My comment: This situation is a vivid symptom of the crisis of trust in the age of AI. When contractors save on expertise by offloading work to neural networks, it is not only the reputation of a specific company that suffers, but also the quality of government decisions. The industry urgently needs standards for source verification and strict sanctions for their violation, otherwise we risk building policy on a foundation of fabricated data.