The Report That Cited a Book Nobody Wrote

Dr Christopher Rudge was reading a government report on welfare compliance systems when a footnote stopped him cold. It cited a book by his University of Sydney colleague Lisa Burton Crawford. Crawford is a professor of public and constitutional law. The book the footnote described was about something else entirely, in a field she has never worked in.

Rudge told the Associated Press exactly what went through his head: "I instantaneously knew it was either hallucinated by AI or the world's best kept secret because I'd never heard of the book and it sounded preposterous." He'd stumbled onto the world's best-kept secret roughly as often as anyone does. Which is never. It was hallucinated.

The report attached to that footnote cost the Australian government up to $440,000, depending on which outlet you believe, and the fact that even the price tag is contested tells you something about how this whole affair went.

What the report was actually for

The Department of Employment and Workplace Relations commissioned the report in December 2024. Its job was to review the targeted compliance framework, the IT system that automates penalties inside Australia's welfare system. This is a high-stakes corner of government. This is exactly the kind of report that needs its citations to survive contact with a lawyer.

Deloitte's Australian arm delivered 237 pages. It went up on the DEWR website on 4 July 2025. Somewhere in those pages, alongside real findings about a lack of "traceability" between the framework's rules and its underlying legislation and genuine "system defects," sat references to research that had never been written.

The mechanism: fluent and wrong

The revised version of the report later disclosed what had gone into producing it: "a generative artificial intelligence (AI) large language model ... based tool chain licensed by DEWR and hosted on DEWR's Azure tenancy." In plain terms, someone used a chatbot to help write a government compliance review, and the chatbot did what chatbots do when you ask them for a citation. It produced something citation-shaped.

A large language model predicts the next plausible word. Ask it for an academic reference on welfare law and it will generate a reference that looks exactly like the ones it was trained on: a plausible author and title, with a publication year that happens to fit. Whether the book exists is nobody's job inside the machine. It answered anyway, with the same fluent confidence it uses for everything else, and nobody downstream checked before the file went to a federal department.

The AFR's count came to three nonexistent academic references in the original version, including phantom research attributed to professors at the University of Sydney and Lund University in Sweden. And then there was the one that should have set off every alarm in the building: a fabricated quote, attributed to a real Federal Court judgment.

The quote that no judge ever wrote

The invented citation landed in a robodebt-adjacent case, Deanna Amato v Commonwealth.

A litigator catches a fake quote from a sitting judgment, because sooner or later someone reads the actual judgment and finds nothing resembling the words on the page. Deloitte later amended that reference. It stayed in the report long enough to be published on a government website first.

The quiet fix

The AFR broke the story in August 2025, citing Rudge's findings. Once the story was public, Deloitte skipped the press conference. A revised version of the report went up on the DEWR website on a Friday, ahead of a long weekend across much of Australia. It was dated 26 September and replaced the July version without much fanfare. If you'd bookmarked the original PDF, you'd have no idea it changed. If you'd cited it in your own work already, that's your problem now.

The department's line was that Deloitte had "confirmed some footnotes and references were incorrect." Deloitte's line was that "the updates made in no way impact or affect the substantive content, findings and recommendations in the report." A spokesperson added that "the matter has been resolved directly with the client." Three sentences of institutional Teflon, each one true in the narrowest possible sense and doing a lot of work to avoid the word "hallucination."

The refund, and the argument about the refund

Deloitte agreed to a partial refund. Fortune and the Associated Press put the contract at $290,000. The AFR, Guardian, Financial Times and news. com. au all reported $440,000. Nobody in this story seems entirely sure what the government paid for a report that turned out to need an errata sheet, which is its own small joke.

The partial refund left people cold. Greens senator Barbara Pocock said Deloitte "misused AI and used it very inappropriately: misquoted a judge, used references that are non-existent," and argued for a full refund rather than a partial one. Labor senator Deborah O'Neill was blunter: "Deloitte has a human intelligence problem. This would be laughable if it wasn't so lamentable. A partial refund looks like a partial apology for substandard work." A partial refund for a report that mixed real findings with invented sources is, in fairness, a fairly precise metaphor for itself.

The timing nobody could have scripted

Here's the detail that makes the story worth remembering past this news cycle. In September 2025, Deloitte announced it would invest $3 billion in generative AI development through fiscal year 2030. Around the same period, one of the leading AI developers announced a partnership with Deloitte making its models available to more than 470,000 Deloitte professionals. A firm mid-scandal over an AI-generated fake bibliography was, at the exact same moment, rolling out AI access to nearly half a million of its own staff. Somewhere a risk-management department is having a very long month.

The 2am version

Nobody at Deloitte set out to invent a book. Somebody needed citations, a tool produced citations that looked exactly like the real thing, and the gap between "looks real" and "is real" didn't get checked before 237 pages went to a federal department. The findings underneath may well have been sound. The sourcing was fiction, and sourcing is the part a reader can actually verify, which is exactly why it's the part that sank the story.

This is a citation problem, and citation problems don't check your job title before they happen. A student pulling references for a thesis at 2am is running the same risk as a $440,000 engagement team: a tool that writes a source-shaped sentence with total confidence and zero obligation to have it be true. Rudge caught his fake book because it happened to land in his own field, on his own colleague's name. Most fabricated references don't get that lucky. They get filed or published, and the gap surfaces later, in someone else's inbox, at a worse possible time.

The boring fix is the same size as the problem: check every citation against the live record before you submit anything with your name on it. Does the paper exist. Does it say what you're claiming it says. Is it still standing, or has it been retracted since the version you found. All of that requires trusting nothing until you've looked, which is a policy Deloitte's report writers might want to adopt before the next long weekend.