Key facts at a glance
- Western Australia’s government is seeking a fast refund from Deloitte over a flawed cost-benefit analysis about merging three public universities.
- The state paid A$166,000 for the report, which examined a proposed merger of the University of Western Australia, Curtin University, and Murdoch University.
- The report contained AI-generated references to sources that do not exist, as well as incorrect cost totals in its appendix.
- Deloitte said it used generative AI to help gather research, summarise information, and proofread the document.
- The government says it has been told the errors do not change the report’s modelling or findings, and Deloitte will correct and reissue it.
- The case follows a similar 2025 episode in which Deloitte agreed to partly refund the federal government for an A$440,000 report that also contained AI-generated errors.
Western Australia’s government is pressing Deloitte for a swift refund after a publicly funded report on merging three of the state’s universities was found to contain fabricated references and incorrect cost figures. The state paid A$166,000 for the cost-benefit analysis, which was meant to inform a high-stakes decision about the future of the University of Western Australia, Curtin University, and Murdoch University. Instead, the document has become a case study in the risks of using generative AI in professional advice, particularly when public money and major policy decisions are involved.
A government spokesperson said the state was very disappointed and had written to Deloitte to speed up the refund. The wording is notable: the government is not merely asking for corrections, but for money back. That signals how seriously the state views the combination of invented citations and arithmetic errors in a report that was supposed to provide a reliable evidence base. Deloitte has not responded to questions about the matter.
A report meant to guide a major university merger
The report was a cost-benefit analysis of a proposal to merge the University of Western Australia, Curtin University, and Murdoch University. University mergers are complex public-policy exercises. They involve not only balance-sheet questions about operating costs and capital spending, but also academic workforce planning, student access, research capacity, regional impact, and institutional identity. A cost-benefit analysis in that setting is expected to be transparent, traceable, and defensible. Every figure should be sourced; every assumption should be testable; every citation should point to a real document.
According to the available account, Deloitte said in the report that it used generative AI to help gather research, summarise information, and proofread. Those are common uses of AI in professional services, and they can improve speed. But they also create specific risks. AI systems can produce plausible-sounding but non-existent references, a problem often described as hallucination. They can also introduce subtle numerical errors when material is summarised, copied between sections, or reformatted. In a report with an appendix of cost estimates, even a small transcription mistake can change the apparent scale of savings or costs.
Fabricated references and wrong totals
Critics first flagged the made-up references earlier in October. The concerns began with an online post and then spread among people scrutinising the document. At the time, Deloitte characterised the problems as minor errors, and the government instructed the firm to fix and reissue the report. That response suggested the issue might be contained: correct the citations, check the text, and republish.
But further scrutiny found problems beyond the references. Sums in the report’s appendix were wrong. In one example, two costs for a Curtin and UWA merger totalled A$659 million, but the report gave A$399 million. The gap is not a rounding difference. It is a substantial discrepancy in a document whose purpose is to weigh costs and benefits. Even if the error sits in an appendix rather than in the main narrative, appendices are part of the evidentiary record. Decision-makers, journalists, stakeholders, and opponents of a merger can all rely on them.
The government said it had been told the errors do not change the report’s modelling or findings. Deloitte will correct them and reissue the report, according to the spokesperson. That may be true, but the public is entitled to ask how the errors occurred, how they were missed, and what quality controls were in place before the report was delivered. The refund request adds a commercial dimension: the state is testing whether a consultant should bear financial responsibility when a deliverable fails basic accuracy checks.
Why the refund demand matters
The demand for a fast refund is significant because government procurement usually moves slowly. Contracts can include review periods, remediation clauses, and dispute-resolution steps. A public call to speed up a refund suggests the government wants to signal that the errors are not being treated as a technicality. It also puts pressure on Deloitte to resolve the matter quickly rather than allow it to become a prolonged contract dispute.
For Deloitte, the reputational stakes are clear. The firm is one of the largest professional services organisations in the world, and its advice is used by governments and companies for decisions involving large sums. A flawed report on university mergers in Western Australia is a local issue, but it raises global questions about how consulting firms supervise AI-assisted work. If generative AI is used to gather research and proofread, firms need robust verification. Citations should be checked against original sources. Numbers should be reconciled independently. Drafts should be reviewed by people with the expertise to spot implausible figures.
The case also highlights an accountability gap. When a report contains fabricated references, it can be difficult for a client to detect the problem without redoing the research. Governments often lack the in-house capacity to audit every citation in a consultant’s report. They rely on the consultant’s professional standards. That reliance is reasonable, but it also means that failures can damage public trust far beyond the value of the contract. A A$166,000 report can shape a debate worth hundreds of millions of dollars.
Not an isolated incident
It is not the first time Deloitte has faced questions over AI-generated errors in a government report. In 2025, the firm agreed to partly refund Australia’s federal government for an A$440,000 report that also contained AI-generated errors. The parallel is uncomfortable. It suggests that the problem is not a one-off slip by an individual author, but a systemic risk in how AI tools are being adopted in high-value advisory work.
Rival firms KPMG and PwC have faced similar problems with AI-written reports this year. The pattern matters because the consulting industry is under pressure to demonstrate that it can use AI responsibly. Clients want efficiency, but they also need accuracy. Governments want modern tools, but they cannot outsource accountability to a model. The more AI is used to draft, summarise, and check documents, the more important it becomes to have human verification that is independent, documented, and proportionate to the risk.
What happens next
The immediate next steps are straightforward on paper. Deloitte is expected to correct the errors and reissue the report. The Western Australian government wants a refund quickly. The state says it has been assured that the errors do not alter the modelling or findings. If that assurance holds after reissue, the policy debate over the university merger can continue on the basis of corrected evidence. If it does not, the consequences could be more serious.
There are also broader questions. Will the corrected report disclose exactly what changed? Will the government publish the revised version alongside the original? Will the refund cover the full A$166,000 or only part of it? Will there be an external review of how the AI tools were used? Those questions are not answered in the current public account, but they are likely to shape how the episode is remembered.
For now, the core facts are clear. Western Australia paid A$166,000 for a report on merging three universities. The report contained AI-generated references to sources that do not exist. It also contained incorrect cost totals, including an appendix figure of A$399 million where two costs totalled A$659 million. Deloitte says it will correct and reissue the report. The government is seeking a fast refund. And the episode has become part of a wider pattern of AI-related errors in consulting reports, adding pressure on firms to prove that speed has not come at the expense of accuracy.