AI Hallucinations: Australian Firms Own the Mistake, Not the AI

Deloitte refunded more than $97,000 to the federal government last year after a report it was paid about $440,000 to produce turned out to be salted with things that never existed — a quotation attributed to a Federal Court judgment that was never spoken, and references to academic papers that were never published. The Department of Employment and Workplace Relations had commissioned a 237-page review of its welfare compliance system. Part of what it received was AI output that nobody had checked. A revised version quietly disclosed that Azure OpenAI had been used in the drafting.

That is the cleanest illustration yet of a rule every Australian professional-services firm now operates under: when AI invents a fact and it goes out under your name, the firm wears it — not the software, not the vendor. Deloitte is a global firm with a legal department and professional indemnity cover most practices can only envy. It still paid the money back, and it still took the reputational hit on every front page in the country.

The pattern isn't confined to consulting. In August 2025 a Victorian solicitor became the first lawyer in Australia sanctioned over AI: he had filed a list of case citations, generated by AI, that pointed to decisions which did not exist. The Victorian Legal Services Board varied his practising certificate — he can no longer practise as a principal, run his own firm, or handle trust money. Weeks earlier, a King's Counsel defending a murder trial in the Supreme Court of Victoria took "full responsibility" for submissions containing fabricated quotes and nonexistent citations, and apologised to the bench.

Legal-sector reporting now counts more than twenty Australian court matters since 2024 involving AI-fabricated references. The through-line in every one is the same: relying on AI output without checking it does not meet a professional's obligations, and it makes no difference which model produced the error. Australia's legal bodies spelled the duty out in a joint statement — a practitioner must be able to personally verify anything AI produces before it leaves the office. Swap "court" for "client" and "practitioner" for "accountant, bookkeeper, adviser or consultant", and the standard is identical.

Want to know what this means for your business? We'll map it in one call.

Here's the uncomfortable part. The efficiency is real — AI genuinely drafts the file note, the advice letter, the compliance summary in a fraction of the time. That's admin leverage worth having. But the same tool that saves you two hours will, with total confidence, cite a superannuation ruling that doesn't exist or misstate a client's tax position, and it will do it in prose so clean that a busy reviewer nods it through. The risk isn't that AI is wrong occasionally. It's that it's wrong invisibly, inside work that looks finished.

For a professional-services firm the exposure is direct. A hallucinated figure in advice that a client acts on is a professional-negligence question, and your indemnity insurer will ask what verification process you had. We've written about who is on the hook when an AI tool misleads a customer under Australian Consumer Law, and the answer there is the same as here: the business that put its name to the output. "The AI got it wrong" is not a defence a court, a regulator or a client has shown any patience for.

If you're not sure where AI already sits inside your firm's work product, that's the first thing worth mapping — see where your firm is most exposed before you tighten anything.

You don't solve this by banning AI — the firms that do just lose the productivity and drive the usage underground. You solve it with a verification gate: one documented step where a named person checks every AI-assisted output against source before it reaches a client or a court. Every citation traced to a real authority. Every figure reconciled to a real record. It sounds obvious; almost nobody has written it down.

Start with the highest-stakes work product — advice, filings, compliance reports — and make the sign-off explicit and attributable. From 10 December there's a second reason to have this in order: new Privacy Act rules mean you may also have to tell customers when AI shapes a decision about them, and a firm that already knows where AI sits in its workflow is halfway to compliant.

Key takeaways

Common questions

Related notes

Do you know where AI already sits in your firm's work?

Most firms can't say which of their outputs are AI-assisted, let alone who checks them. Book a call and we'll map your AI exposure and design a verification gate that keeps the speed without the liability.