The A$440,000 hallucination: what the Deloitte refund tells us about AI in professional services

Deloitte has agreed to partially refund the Australian government after a report it was paid A$440,000 to produce turned out to be riddled with mistakes the firm has now admitted were generated with the help of AI.

The report was commissioned by the Department of Employment and Workplace Relations (DEWR) and published in July 2025. Within weeks, academics reading it started finding things that could not be true: footnotes citing papers that did not exist, quotes attributed to judges who had never said them, and paragraph numbers pointing to parts of court rulings that were simply not there.

The common thread, eventually confirmed by Deloitte, was that a generative AI tool (Azure OpenAI GPT-4o) had been used in the report's drafting.

What a fabricated footnote looks like

The first person to pull on the thread was Dr Chris Rudge, a welfare academic at the University of Sydney. He recognised his colleagues' names in the citations and immediately spotted that the works attributed to them did not exist.

"I was in no doubt, when I read the names of the works, that they were fake because I work in the area and I know my colleagues' work," Rudge told the Australian Financial Review. "Once I discovered one, I just discovered more and more."

Among the fabrications:

• A paper titled The Rule of Law and Administrative Justice in the Australian Social Security System. It does not exist.
• An article called Choice, Responsibility and the Regulation of Behaviour: Lessons from the Social Security System, attributed to Carolyn Adams and the UNSW Law Journal, 2012. Also a fiction.
• A quote attributed to 'Justice Davis' (a misspelling of Justice Jennifer Davies) in the Deanna Amato v Commonwealth robo-debt case, citing paragraphs 25 and 26 of the consent orders. Those paragraphs do not exist. Another passage was attributed to paragraph 30 but actually appeared in paragraph 9.

"I cannot understand how a human could create titles of works that don't exist, that don't appear on Google," Rudge said. "I think it's good if it's AI, because to think of a person doing that is almost worse. It's very disrespectful to those who have done the research to just not get it right."

The most Australian AI scandal possible

Strip away the professional politeness and the story is this: a Big Four consultancy took public money to produce an independent assurance review of a welfare compliance system, and handed in work that contained invented evidence about the very judgment, Amato, that helped expose the robo-debt scandal in the first place.

Robo-debt was the automated decision-making failure that wrongly pursued vulnerable Australians for debts they did not owe. It cost lives. The Royal Commission's final report made plain how badly a government can hurt its own citizens when it trusts automation it does not actually understand.

Against that backdrop, Deloitte produced a review of another welfare compliance framework, and let an AI hallucinate quotes from the very court ruling that helped end the last disaster.

That is not a clerical error. It is a governance failure with a particular historical sting.

'The substance of the review is retained'

Deloitte's response has the familiar shape of a firm trying to keep the client relationship intact.

A corrected version of the report was re-uploaded. Then uploaded again, this time with a note acknowledging 'a small number of corrections to references and footnotes'. A DEWR spokesperson confirmed that 'some footnotes and references were incorrect', but added: 'The substance of the independent review is retained, and there are no changes to the recommendations.'

On Deloitte's part, a spokesperson said 'the matter has been resolved directly with the client.'

Read that sequence carefully. The story is being framed as a footnote problem. The citations were wrong. The recommendations were not. Move on.

The steelman: AI-assisted drafting, corrected and refunded

The charitable read of Deloitte's position is the version most of the firm's peers will reach for when their own version of this lands.

It runs like this. Generative AI is now a normal tool in long-form professional drafting. Lawyers use it. Auditors use it. Accountants use it. The recommendations of the DEWR review were derived independently by experienced consultants and informed by their own subject-matter judgment. The citation errors are real, and embarrassing, but they are correctable: the file has been re-uploaded, the references fixed, the client made whole through a partial refund. No reader was misled into a different policy conclusion. The substance stands.

That is not a stupid argument. A firm that owns its mistake, fixes the artefact, and returns part of the fee has done more than most.

The trouble is that the argument depends on a split the evidence professions do not actually recognise. In a court, a judgment that rested on fabricated authorities would be vacated, not corrected with a footnote erratum. In academia, a paper with invented citations is retracted, not re-uploaded. The substance of an evidentiary document cannot be cleanly separated from the evidence it cites, because the cites are how the reader is invited to check the substance. A review of a welfare compliance system that misquotes the leading robo-debt judgment is not a sound review with a small reference problem. Its evidentiary scaffolding has been shown to be unreliable, and its recommendations now sit on top of work no one has fully verified.

A partial refund settles the contract. It does not answer the harder question, which is whether the work was ever what the client thought they were buying.

The AI Institute problem

There is one detail that makes this episode difficult to dismiss as a quirk.

Deloitte sells AI advisory services. The same firm that failed to catch an AI hallucination in its own deliverable runs an AI Institute advising other organisations on how to adopt AI responsibly. Its consultants charge premium rates to tell clients how to govern exactly the kind of workflow that just embarrassed them.

If the firm cannot run the review process reliably on its own flagship public-sector engagement, what exactly is it selling when it promises to stand up governance for someone else?

This is not a gotcha. It is the central question. The value proposition of a Big Four consulting firm is that it brings discipline, expertise, and review. When those controls fail on a A$440,000 report, the product being sold is not the report. It is the process. And the process failed.

Where were the humans?

A generative model will hallucinate citations. That is a known property of the technology, not a surprise. Anyone who has spent twenty minutes with GPT-4o knows that it will happily invent a plausible-sounding academic paper and attribute it to a real scholar.

The failure at Deloitte was not that an AI produced false citations. It was that no human checked them before the report went out the door.

That is the governance gap worth staring at. Someone wrote the first draft with AI assistance. Someone edited it. Someone approved it. Someone signed it off to a government client. At none of those steps did anyone open a browser, search a footnote, and confirm that the cited paper was real.

In the professional services model, the partner is the human in the loop. That is what the client is paying for: the judgment and the signature of an experienced professional staking their reputation on the work. On this engagement, that check did not happen.

The economics of the shortcut

Why would it not happen? Because the incentives push the other way.

AI drafting compresses the hours needed to produce a long-form report. That expands the margin on fixed-fee engagements, or lets the same team service more clients. The shortcut is financially attractive at every layer: for the junior who drafted it, for the manager who reviewed it, for the partner who signed it. The one party not materially served by the shortcut is the client, who is paying for human diligence and receiving machine output.

This is how ethical failure compounds quietly. No one decides to mislead the government. A series of small efficiency choices adds up to a report no human has verified end-to-end. The AI fills the gap left by the missing review step, confidently, fluently, and wrong.

Who loses

It is easy to read the Deloitte headline as a wonkish footnote story. It is not. The cost of this engagement was paid in specific places by specific people.

Australian taxpayers. A$440,000 of public money was spent on an independent assurance review of a welfare compliance system. A partial refund returns some of that money, but not the time of the public servants who briefed the work, processed the deliverable, and now have to decide what to do with a report whose footnotes have been shown to be unreliable. The recommendations may stand. The trust in them does not, and the cost of re-establishing it falls on the department.

Academics whose names were attached to invented papers. Carolyn Adams and the colleagues Rudge recognised did not write the works the report attributed to them. Their names now sit, in a Commonwealth document, against research they never produced. As Rudge put it: 'It's very disrespectful to those who have done the research to just not get it right.' For working academics, citation is currency. A government report that mints fake currency in their names is a small violence done quietly.

Robo-debt survivors. The Amato case is one of the legal landmarks that helped expose the robo-debt scandal. Misquoting that judgment in a Commonwealth report on a successor welfare compliance framework is not a neutral error. It tells the people the system has already harmed once that the institutions reviewing the next system are not reading the last one carefully. That is the part that stings, and it is invisible in the refund figure.

Future Deloitte public-sector clients. Every report the firm hands in from here will be read with the question: did anyone open the footnotes? That is corrosive for the public servants whose job is to commission this work and defend the spend, and who now carry the verification burden the firm was paid to discharge.

What responsible AI use looks like in consulting

Using generative AI to accelerate drafting is not the problem. The problem is using it without a verification layer.

Responsible AI use in professional services needs three things the pre-AI consulting world already knew how to do, now reapplied to machine-generated text.

1. Source checking. Every citation in an AI-assisted draft is treated as unverified until a human has opened the original source. No exceptions. The cost of this check is trivial compared to the cost of getting it wrong.
2. Disclosure to the client. If an AI tool was used in the drafting, the client is told. DEWR was only told after academic readers caught the errors. That is disclosure in the wrong direction.
3. Accountability for the output, not the input. 'The AI wrote it' is not a defence. The firm signs the report. The firm owns the content, including the made-up footnotes.

None of this is novel. It is how evidence has always been handled in law, in academia, in audit. What is new is that a particular class of errors, invented but plausible citations, can now be produced in bulk by a tool that does not know it is lying.

The canary, not the catastrophe

Deloitte is one engagement. A$440,000 is, in the scale of Commonwealth procurement, a rounding error. If this story stopped at the refund, it would be a curiosity.

It does not stop there, because the same workflow is now running, today, inside thousands of professional deliverables no journalist has audited and no academic has happened to read. Law firms are using generative AI to draft submissions. Auditors are using it to summarise evidence. Strategy consultants are using it to assemble the very kind of long-form report DEWR commissioned. In 2023, US lawyers filed court submissions with fabricated case citations after using ChatGPT. Legal and medical journals have retracted AI-assisted papers for similar reasons. The pattern repeats: a tool is slotted into an existing workflow, the human verification step is quietly skipped, the errors look plausible enough to survive a casual read, and they are caught only when a domain expert reads the footnotes.

Most will not. Most reports are read for their recommendations, summarised in a covering email, and filed. Dr Chris Rudge happened to be a welfare academic who happened to know the cited authors. Take him out of the story and the fabricated citations sit on a Commonwealth website, uncorrected, indefinitely.

That is what makes Deloitte the canary rather than the catastrophe. The catastrophe is the version of this story that does not get caught.

The Ethical pillar: whose life gets easier, and whose gets harder

Ethics in AI is not an add-on or a review gate. It is a question about whose life gets easier and whose gets harder as a result of the system. The DEWR engagement is a clean example of the asymmetry that question is built to expose.

Whose life got easier? The team drafting the report. The margin on a fixed-fee engagement. The hours pushed through a tool that does not need a coffee break. AI-assisted drafting did exactly what it promises to do: it compressed the labour cost of producing a long-form professional document.

Whose life got harder? The academics whose names now sit, in a government file, against research they never wrote. The robo-debt survivors who watched the Amato judgment misquoted in a Commonwealth report on the next welfare compliance system. The DEWR officials who now have to defend a deliverable whose evidentiary base has been shown to be unreliable. The taxpayers who paid for human diligence and received machine output. None of them were in the room when the decision was made to draft with AI and skip the verification step. The trade was made on their behalf.

A partial refund returns some money. It does not return the asymmetry. That is the work the Ethical pillar exists to do, and it is the work the next consulting firm publishing an AI-assisted report still has to choose to do, or not.

---

Sources: Cyber Daily / Accounting Times reporting by Daniel Croft (9 October 2025), drawing on Australian Financial Review coverage by Paul Karp and commentary from Dr Chris Rudge, University of Sydney.