Photo_J / shutterstock.com
When the report hallucinates: How artificial intelligence became an embarrassing act
A consulting firm like Deloitte is supposed to deliver quality - especially when the Australian government puts almost half a million dollars on the table for it. But instead there were apparently AI fantasies, false citations and incorrect footnotes. Now some of the money is back.
What had happened? The government under Prime Minister Albanese had commissioned Deloitte to take a close look at the Ministry of Labor's control system - more precisely: an IT system that automatically punishes welfare recipients if they violate their obligations. The report, published in July 2025, was intended to uncover irregularities. And Deloitte found them - such as a lack of traceability between the law and its application as well as "systemic defects".
But it didn't take long for independent experts to notice: Something is wrong here.
AI instead of plain text: the report was a mix of facts and fantasy
The scandal broke when Dr. Christopher Rudge from the University of Sydney took a closer look. He discovered invented sources, false quotations and "hallucinations" - typical errors of generative AI, where it invents things because it doesn't know a real answer.
According to Rudge, even non-existent court rulings and academic papers were cited. This was corrected in the revised version of the report - but not everything. Some errors have been replaced by new ones: "Instead of one invented quote, there are now five to eight others," says Rudge dryly.
In the updated version, Deloitte at least admits that an AI tool was used - specifically a model called GPT-4o from Azure OpenAI. However, they do not want to officially confirm a connection between the errors and the AI.
Repayment yes - responsibility? No way!
However, Deloitte is not returning the money voluntarily: only the last tranche of the 440,000 dollar contract will be refunded. Exactly how much is yet to be announced. According to the government, the content of the report's recommendations will nevertheless remain valid - despite all the mishaps.
Nevertheless, the criticism was not long in coming. Australian Senator Deborah O'Neill spoke of a "human intelligence problem" at Deloitte. Her biting comment: "Maybe a ChatGPT subscription would have been cheaper."
If the auditing body is missing
If you use AI as a ghostwriter for sensitive government analyses, you shouldn't be surprised about fantasy sources and misquotes. The real scandal is not the use of GPT-4o - but the fact that no one checked whether the content was correct.
Half a million dollars for a report that wouldn't even pass as a university term paper? A partial refund is almost merciful. Perhaps this is the future of consulting: instead of expertise, there is generated half-knowledge - and in the end, no one is liable. AI is no substitute for real expertise. Those who blindly rely on machines don't save money - they lose credibility.
