Australia spent 3.48 million Australian dollars to commission an age verification technology test, but ended up in an embarrassing situation due to a report that had been watered down with AI. The British organization ACCS, which conducted the test, initially denied using artificial intelligence, but later admitted to using AI to rewrite parts of the report after The Guardian discovered ChatGPT metadata in the report link.
Fake citations come to light
This key report, which paved the way for a social media ban for Australians under 16, was based on real technical tests. However, the references in the "emerging technologies" section were full of errors. The Guardian found that there were six incorrect references in the same section: some DOIs pointed to non-existent papers, others linked to wrong documents, and the combination of authors, journals, and years did not match any known research. Even more absurdly, the "correction list" submitted by ACCS later also had problems with incorrect years and authors. One of the papers' main authors clearly stated that they had never participated in this study, and the paper was published after the time they claimed to have accessed it.
Government trust further damaged
ACCS insisted that AI was only used to make the language more concise and that the links themselves were correct, so it wasn't an issue. However, the government contractor has a history of using AI to water down reports. Last year, Deloitte returned 440,000 Australian dollars in contract payments due to AI hallucination errors, and the Australian federal government already has a negative attitude toward such behavior. The academic community warned that false citations, regardless of whether they were caused by AI or not, could mislead government decisions and erode public trust. Some lawmakers have already demanded that the contractors apologize and refund the money. The excuse that only 26 pages of a thousand-page report involved citations and revealed "a few errors" is unlikely to ease this trust crisis surrounding AI ghostwriting.
