A 237-page Deloitte report to the Australian government cited 20 fabricated references: phantom academics, invented co-authors, and a direct quote from a federal judge who never said it. The fix is architectural, not editorial.
In July 2025, Deloitte delivered a 237-page report to the Australian Department of Employment and Workplace Relations. The contract was worth roughly AU$440,000 (about US$290,000). The brief: review the government's automated welfare compliance system and recommend improvements. The report looked authoritative: citations to academic papers, quotes from federal court rulings, references to expert opinion. Then a researcher at the University of Sydney sat down to read it carefully, and the report quietly fell apart. By October, Deloitte was refunding the final installment and quietly adding a disclosure that "a generative AI language system, Azure OpenAI, was used in its creation."
Dr Chris Rudge, a health and welfare law researcher at Sydney, was flagged to the report by a colleague who thought one of the cited cases sounded suspicious. He pulled the citation. The case didn't exist. He pulled another. Also fake. By the time he was done, he had documented around 20 fabricated or distorted references in the report: phantom academic papers attributed to real researchers, invented co-authors, and, most strikingly, a fabricated direct quotation from a federal court judgment. The quote wasn't a paraphrase. It was a string of words in quotation marks, attributed to a judge, that the judge had never said.
Deloitte issued a revised version of the report. The revised version is, by all accounts, fine: most of the substantive recommendations stand. What changed is that the citations are now real, the quotes are now from documents that exist, and the report acknowledges AI was involved in its production. Deloitte agreed to refund the final payment installment to the government as part of the settlement.
The fallout went beyond one report. The UK Financial Reporting Council had already warned in June 2025 that the Big Four were failing to adequately monitor how AI was affecting audit quality. The Australian case turned that warning into a concrete data point: AI-assisted professional services work was shipping to clients with fabricated content embedded in it, and the clients were not catching it.
The Deloitte report was almost certainly produced by a workflow that looks something like this: an analyst writes a draft, then uses a model like GPT-4o or Claude to "strengthen the evidence base," to find more academic backing for an existing claim, or a court ruling that supports a point. The model obliges, returning a paragraph that says exactly what the analyst was hoping for, with citations attached. The analyst pastes it in.
This is the most insidious form of hallucination because it confirms what the user already wants to believe. A skeptic might double-check a surprising citation. Nobody double-checks a citation that supports their existing argument. And the model, optimized to be helpful, will always find something, even if "finding" means generating it.
The same dynamic plays out in AI analytics. A marketer asks: "Did our paid social spend pay off last quarter?" If the model has a habit of producing satisfying answers, it will produce a satisfying answer. Whether the supporting numbers exist in the warehouse is a separate, easily-skipped question. Confirmation bias is the user's contribution; fluent confabulation is the model's. Without architectural guardrails, the combination is silent and toxic.
Askalytics' AI queries your own data directly, and every number it reports traces back to the exact query that produced it. A second AI re-checks the rest.
The structural fix is to make every number traceable to the query that produced it, and to have a second model re-check the conclusion. A claim you can open and verify is a claim that cannot quietly be invented.
Askalytics' AI Insights pipeline is built on this. The AI queries your own data warehouse directly and runs the queries it needs to answer the question. The difference is that every number it reports traces back to the exact query that produced it. You can open the receipt for any figure, the Advanced Breakdown, and see the query, the rows, and the source it came from. Run the same query against the same data and you get the same number, every time. On top of that, every AI Insight is re-checked by a second, different AI model, so a confident answer the data does not actually support gets flagged before it reaches you.
This is the same principle that turns scientific writing into a citation-driven discipline. Researchers don't sit down to write a paper and then invent the experiments to support it; they run the experiments first, then write a paper grounded in what those experiments produced, and others can reproduce the result. The Deloitte report failed because it inverted the order: it wrote first, then generated supporting citations after, and nobody could reproduce them. Askalytics ties every figure to a query you can open and read, and puts a second AI in the path before any insight ships.
The Deloitte case isn't really about Deloitte. It's a public, costly, fully-documented example of what happens when professional-grade work is delivered through a workflow that doesn't enforce evidence provenance. The same workflow ships AI analytics insights to growth teams and CFOs every day. Some get caught and refunded. Most don't.
For more on how Askalytics designs around this, read about our evidence-grounded insights pipeline and the Confidence Check that re-checks every insight.