Deloitte billed the Australian government A$439,142 for an assurance report citing books that don't exist

The assurance the department paid for

In December 2024 the Australian Department of Employment and Workplace Relations paid Deloitte to check work the department could not vouch for on its own. It commissioned an independent assurance review of the Targeted Compliance Framework, the automated system the department uses to enforce welfare recipients' obligations.¹ That framework sits in the same corner of welfare administration as robodebt, the debt-recovery program a royal commission later found unlawful. An assurance review is a paid, arm's-length examination, and its whole value is that someone outside the department reads the system and puts a firm's name to the finding.

The contract was worth A$439,142.00.² Deloitte carried out the review between December 2024 and June 2025, dated its report 4 July 2025, and the department published it on 14 August.¹

A book cited repeatedly that was never written

The footnotes did not survive a close reading. In late August 2025, Christopher Rudge, a lecturer in law at the University of Sydney's law school,³ went through the report and found that a number of its authorities did not exist. He counted close to twenty errors, among them roughly ten citations to a single book that was never written, and took what he found to the Australian Financial Review, which broke the story.⁴

Two of the people named are real, and their absent works are not. Lisa Burton Crawford is a professor of public and constitutional law at the University of Sydney Law School. She did not write the book cited repeatedly against her name.⁷ Björn Regnell is a professor of software engineering at Lund University in Sweden. The release-planning paper the report attributed to him was not his.⁸

Worse than any missing book was a court that never spoke. The report's treatment of IT-system defects turned on Amato v Commonwealth, a Federal Court matter from the robodebt line of litigation, which it cited as Amato v Commonwealth [2021] FCA 1019 and quoted directly.

There is no 2021 decision. Amato was resolved in 2019, by consent orders that Justice Davies, not Davis, made in the Federal Court on 27 November. Victoria Legal Aid, which ran the case, records what the court's annexed statement actually said: that "there was no material before the decision-maker capable of supporting the conclusion that a debt had arisen under the Act and, therefore, the conclusion was irrational."⁹ The report's version paraphrases that gist and attaches it to a judge's name, a set of paragraph numbers, and a citation that were all wrong.

Real academics had their names on books and papers they never wrote. A Federal Court judge was quoted saying words she never said. All of it sat inside a document the Commonwealth had paid a global firm to get right.

Why a model invents a citation that looks real

How does a report cite a book that was never written? The answer was in the report itself, if a reader reached page 147. Deloitte disclosed, in two places, that its technical review had used a generative AI tool: an Azure OpenAI GPT-4o system, licensed by the department and run on the department's own servers. Neither disclosure appeared on a cover page or in a summary. One sat inside a methodology graphic on page 48. The other was in Appendix C, "Detailed Methodology: Technical Review", a note that the work "included the use of a generative artificial intelligence (AI) large language model (Azure OpenAI GPT-4o) based tool chain licensed by DEWR and hosted on DEWR's Azure tenancy."⁶

A model of this kind does not look a reference up and copy it down. When it produces a citation, it generates the most statistically likely run of words to follow the ones before, the same way it produces any other sentence. A real reference and an invented one come out of the identical process, so a fabricated citation carries every surface mark of a true one: a plausible author, a plausible title, a plausible year and page. Nothing in the model confirms that the work exists, because confirming is not part of what it does.

The same process has produced the same failure before: a science model that fabricated references by design (Meta's Galactica), and a New York lawyer who filed six invented cases in a real court (the Avianca brief). Why a model builds a fake citation exactly as it builds a true one is the subject of an earlier lesson. What is new here is the setting: not a demo and not one careless filing, but a paid deliverable a government commissioned, received, and published.

Deloitte corrected the report and refunded a fifth of the fee

On 26 September 2025 Deloitte issued a corrected report, which the department published on 3 October. A "Report Update" note on the second page recorded what changed: citations and reference-list entries "which contained errors", the summary of the Amato proceeding, and revisions "to improve clarity and readability". The same note stated that the updates "in no way impact or affect the substantive content, findings and recommendations in the Report".⁶ A department spokesperson likewise said only that "some footnotes and references were incorrect", with "no changes to the recommendations".¹⁰

Deloitte agreed to repay part of the fee. At a Senate estimates hearing in early October, officials confirmed the amount.

The refund was A$97,587.11, the contract's final instalment, or about a fifth of what the department had paid.¹¹

Four things about the episode stay genuinely contested, and the record settles none of them. The first is scale. The department called the problem "a small number of corrections to references and footnotes";¹ Rudge, who found them, put it near twenty errors, including the roughly ten citations to one non-existent book.⁴ Both can be literally true, since a small number of distinct fabricated works can be cited many times over.

The second is whether the correction was substantive. Deloitte's note insists the findings and recommendations did not move, and on the record they did not.⁶ Critics answer that fabricated legal authorities inside a legal-assurance report bear on whether the report can be trusted at all, whatever survived in the recommendations.⁴

The third is the refund. The government and Deloitte treat repayment of the final instalment as resolution. The Australian Greens, whose senator Barbara Pocock drew the figure out at the hearing, called it "measly" and demanded the full A$439,142 back.¹¹ No one disputes the number. They dispute what it should have been.

The fourth is the cause. The disclosure and the pattern of the errors point to unverified model output entering the citations. Deloitte has published no detailed account of how the fabrications got in, so where exactly the human process failed is asserted by others rather than admitted in the document.⁶ The correction did not close the file either: on 3 February 2026 the department published a further-corrected version, replacing the September one.¹

Where an unchecked citation gets in

What went wrong at Deloitte was a missing step, and the step is missing far beyond Deloitte. A generative AI tool helped produce a paid deliverable,⁶ and no one between the tool and the client opened a single citation to confirm the source was real. The same gap sits wherever these tools now draft: consulting reports, legal filings, and the submissions governments commission and publish. It reached journalism too, when CNET ran AI-drafted finance articles under a staff byline. The tool returns a fluent, correctly formatted reference on demand, and it reads exactly as trustworthy as a real one until someone checks. In each setting, the person who signs the document is the last check on whether its sources exist.