Home › Stories
Deloitte Australia · 2024–2025 · An AI story

The World's Best Kept Secret

Deloitte's A$440,000 review for the Australian government cited work that did not exist and put invented words in a Federal Court judgment. The corrected version disclosed that a GPT-4o tool chain had been used.

◷ 2 min read

Chris Rudge researches health and welfare law at the University of Sydney. He was reading a 237-page government report when he hit a reference to a book by one of his colleagues, a professor of public and constitutional law. He had never heard of it. The title suggested a subject well outside her field.

“I instantaneously knew it was either hallucinated by AI or the world’s best kept secret,” he later told the Associated Press, “because I’d never heard of the book and it sounded preposterous.”

The report had cost the Department of Employment and Workplace Relations about A$440,000. In December 2024 the department had hired Deloitte to review the Targeted Compliance Framework, the IT system that automates penalties for jobseekers on welfare who don’t meet their mutual obligations. The review, published on the department’s website on 4 July 2025, was hard on that system. It called it “driven by punitive assumptions of participant non-compliance.”

Rudge kept reading. By his count there were up to 20 errors. References pointed to reports that did not exist, credited to professors at the University of Sydney and at Lund University in Sweden. His colleagues’ work, he said, had been used as “tokens of legitimacy,” cited but not read.

What bothered him more was a robodebt case heard in the Federal Court, Amato v Commonwealth. The report quoted the judgment. The quote was made up.

“That’s about misstating the law to the Australian government in a report that they rely on,” Rudge said. He alerted the media, and in August the Australian Financial Review reported the errors.

On Friday 3 October, ahead of a long weekend across much of Australia, a new version went up on the department’s website. More than a dozen nonexistent references and footnotes were gone, the reference list had been rewritten, and the quotes attributed to the judge had been removed. An appendix now disclosed something the July version had not: part of the work had used a tool chain built on a large language model, Azure OpenAI GPT-4o, licensed by the department itself and hosted on its own Azure tenancy.

Deloitte did not say AI was the reason for the errors. Asked by AP whether the errors were AI generated, it did not respond. The revised report said the updates “in no way impact or affect the substantive content, findings and recommendations in the report,” and the department confirmed the recommendations had not changed.

Rudge read the new version closely. In the new version, where one fake reference had stood, there were now five, six, seven or eight real ones. “So what that suggests is that the original claim made in the body of the report wasn’t based on any one particular evidentiary source.” He still would not call the whole report illegitimate, because its conclusions matched plenty of other evidence.

Deloitte agreed to repay the final instalment of its contract. Labor senator Deborah O’Neill, who had sat on a Senate inquiry into the integrity of consulting firms, said: “Deloitte has a human intelligence problem.” On 21 October the department said the refund came to more than A$97,000.

The department’s page for the report now carries one more line. On 3 February 2026 it was updated again, “to address identified corrections.”

What this story shows

Human-in-the-LoopThe revised report disclosed that a GPT-4o based tool chain was used in part of the work. The fake references in the first version were found after publication, by an academic outside the project.Segregation of Duties and the Four-Eyes PrincipleThe review was meant to be a second pair of eyes on the department's welfare system. The errors in the review itself were spotted by a reader of the published report, who said colleagues' work had been cited but not read.Cost of QualityFixing the references after publication meant a quietly reposted report, weeks of headlines, a refund of more than A$97,000, and another round of corrections months later.Principal–Agent ProblemThe department paid about A$440,000 for outside expertise and relied on what came back. A senator who had sat on an inquiry into consulting firms said anyone hiring them should ask exactly who is doing the work they pay for.

Sources

  1. Department of Employment and Workplace Relations, Targeted Compliance Framework Assurance Review, Final Report (report page and revision notes)
  2. The Guardian, Deloitte to pay money back to Albanese government after using AI in $440,000 report, October 6, 2025
  3. Associated Press, Deloitte to partially refund Australian government for report with apparent AI-generated errors, October 7, 2025
  4. Australian Financial Review, Deloitte AI report: Firm admits to using AI in error-filled government report, October 5, 2025
  5. CFO Dive, Deloitte refunds over $60K for report with AI errors, Australian government says, October 21, 2025