Western Australia Wants Its Money Back From Deloitte Over a $166,000 University Report With AI Hallucinations and a $311 Million Math Error

The Western Australian government says it is "extremely disappointed" with Deloitte and has written to the consultancy to expedite cost recovery for a university merger feasibility study that turns out to contain AI-hallucinated references and cost totals wrong by hundreds of millions of dollars. The Guardian reported the details on Monday. The state paid Deloitte A$166,000 for the study, which examined merging the University of Western Australia, Curtin University, and Murdoch University.

The arithmetic failures are the kind that a spreadsheet should have caught. In the high-cost scenario for a merger between Curtin and UWA, the report lists "total costs to day 1" as A$514 million and "post day 1" costs over five years as A$145 million. Those two lines sum to A$659 million; the report wrote A$399 million — a A$260 million discrepancy. In the low-cost scenario for a three-university merger, A$396 million plus A$116 million should total A$512 million, but the report wrote A$823 million — an error of A$311 million in the other direction. In the mid-cost scenario, A$421 million and A$122 million were combined into A$533 million instead of A$543 million, a A$10 million gap visible on page 216 of the PDF.

The hallucinations are more distinctive. The Guardian found that one reference to a World Health Organization report carries a ChatGPT metadata tag on its URL, indicating the link was sourced directly from the chatbot. The report also repeatedly refers to Curtin University as "Curtin University of Technology" — a name the institution abandoned in 2010 and formally changed under an act of state parliament in 2017. The report's own conclusions contradict each other: the opening claims a three-way merger has the highest cost-effectiveness ratio, while the data tables show a two-way Murdoch-UWA merger scoring better.

Deloitte admits in the report that it used generative AI to "support the collation of research, synthesis of information and editorial proofing," while insisting it did not use AI-generated content, models, or drafts directly, and that all AI-assisted results were fully checked by professionals. When the fabricated references surfaced earlier this month, the firm described the errors as "minor." The state opposition has taken to calling the document "AI slop." Western Australia's tertiary education minister, Tony Buti, first called the problems "typos," then said a day later he was angry and disappointed with Deloitte.

Governance experts were blunt. Professor Stephen Bartos, a public sector governance specialist at the University of Canberra, said the errors discredited the report and were so extensive that it probably needs to be rebuilt from the ground up without AI. Professor Christian Downie of the Australian National University noted that while human error happens, Deloitte disclosed its AI use only in broad strokes, and governments cannot have confidence in reports where the division of labor between machine and human is opaque.

This is not a first offense for the industry. In 2025, Deloitte agreed to partially refund the Australian federal government for a A$440,000 report that also contained AI-generated errors, and KPMG and PwC have faced similar problems this year. The Western Australian government has been advised the errors do not affect the report's underlying modeling, and Deloitte will correct and reissue the document — but the state has attached something new to the episode: a financial consequence. That is the precedent worth watching. When a consultancy's AI-assisted shortcuts produce a nine-figure arithmetic error, "minor errors" stops being a defense, and proofreading becomes a contract term.