Western Australia seeks recovery after report errors emerge
The Western Australian government is seeking an expedited reimbursement from Deloitte after discovering errors in a feasibility study examining potential mergers among the state’s three major universities. The government paid the consultancy $166,000 for the work on the University of Western Australia, Curtin University and Murdoch University.
The dispute centers on a report that included AI-hallucinated references as well as major discrepancies in its cost calculations. The government said it was “extremely disappointed” by the failures and is pursuing cost recovery from Deloitte.
The study considered three merger options and presented low, mid and high cost scenarios. Its appendix set out estimates for costs incurred up to a proposed first day of a merger and for costs expected in the following five years. In at least two scenarios, the written totals did not match the component figures shown in the report.
Hundreds of millions of dollars separated the figures from the totals
In the report’s high-cost scenario for a merger between Curtin and the University of Western Australia, Deloitte listed $514 million in costs to day one and $145 million in post-day-one costs over five years. Those amounts total $659 million. The report instead gave a total of $399 million, creating a discrepancy of $260 million.
Another error appeared in a low-cost scenario involving Curtin, UWA and Murdoch. The study listed $396 million in costs to day one and $116 million over the subsequent five years. Those figures add to $512 million, but the report stated a total of $823 million—a difference of $311 million.
These are not minor presentational mistakes. The estimates concern a politically consequential proposal that could reshape the organization of three universities. When the basic totals do not align with the report’s own line items, the reliability of the broader analysis comes into question.
AI use intensified scrutiny
The report also drew attention after media coverage found hallucinated references. Deloitte acknowledged in the report that it used generative AI to support its work. The state opposition described the result as “AI slop,” while the government moved to recover its expenditure.
The issue is not simply whether a firm used generative AI. The immediate concern is whether a client received a properly checked analysis, with reliable sources and internally consistent figures. In this case, both invented references and arithmetic discrepancies were identified in work commissioned to inform a high-stakes public-policy decision.
Stephen Bartos, a University of Canberra expert in public-sector governance, said the errors discredited the report. He said the material likely needed to be redone or, at minimum, reworked from the ground up manually without AI.
A setback for merger deliberations
Feasibility studies often frame the financial and operational choices that governments face before deciding whether to proceed with institutional change. The errors in Deloitte’s analysis do not themselves settle whether any merger would be beneficial or affordable. They do mean that the report cannot be treated as a dependable guide without substantial correction and review.
The findings are especially significant because the report evaluated multiple combinations of the three universities and presented scenarios meant to quantify their likely costs. Those projections require confidence in the calculations beneath them. When the stated totals depart from the underlying amounts by hundreds of millions of dollars, policy-makers and the public are left without a sound basis for comparing the options.
The government’s request for reimbursement places the focus on accountability for commissioned advice. Deloitte was paid to provide a feasibility study; the discovered flaws raise questions about its quality controls, verification process and use of tools that can generate plausible but false citations.
The broader lesson for AI-assisted consulting
Generative AI can assist with drafting and research tasks, but the episode illustrates the limits of treating output from such systems as verified analysis. A reference that looks credible may not exist. A document that appears professionally formatted may still contain errors that a basic numerical check would expose.
For public bodies, the lesson is practical. Where external reports shape decisions involving substantial public spending or the future of major institutions, reviewers need to validate sources, calculations and assumptions before relying on the conclusions. That need exists regardless of whether AI tools were involved, but the discovery of hallucinated references makes the verification obligation more visible.
Western Australia’s attempt to recover the $166,000 payment is a direct response to the report’s defects. The larger consequence may be a renewed expectation that consultancies clearly explain how AI-assisted work is checked before it reaches a client—and that the final product can withstand ordinary scrutiny of its facts and arithmetic.
This article is based on reporting by The Guardian. Read the original article.
Originally published on theguardian.com
