WA Government Seeks Deloitte Refund Over AI Hallucinations and $500M+ Report Errors
Western Australia is demanding a cost recovery from Deloitte after a $166,000 university merger report revealed major calculation errors and AI-hallucinated citations.
The Western Australian government is seeking an expedited cost recovery from consulting firm Deloitte following the discovery of significant mathematical errors and artificial intelligence hallucinations in a $166,000 state-funded report. The feasibility study, commissioned to evaluate options for merging the University of Western Australia, Curtin University, and Murdoch University, contained multi-hundred-million-dollar calculation discrepancies alongside hallucinated references generated by large language models.
Arithmetic failures and fake citations
The report assessed three potential merger scenarios categorized by low, mid, and high cost structures, breaking down expenses into initial "day one" costs and subsequent five-year post-merger operations. However, substantial arithmetic errors were uncovered in the report's appendix. In a high-cost scenario between Curtin and UWA, the document listed initial costs at $514 million and post-day-one costs at $145 million. While the actual sum equals $659 million, the report stated the total was $399 million—a $260 million discrepancy. Conversely, in a low-cost scenario across all three institutions, initial costs of $396 million and post-day-one costs of $116 million (summing to $512 million) were mistakenly calculated as $823 million, creating a $311 million overstatement.
Beyond financial discrepancies, the study demonstrated clear indicators of unvetted generative AI output. One reference cited a World Health Organisation document with an embedded ChatGPT metadata tag in the URL, confirming it originated directly from the chatbot. Additionally, Curtin University was repeatedly referred to as the "Curtin University of Technology," an outdated name officially retired by state parliament legislation in 2017. Deloitte acknowledged in the text that generative AI was utilized to assist in research collation, information synthesis, and editorial proofing, though the firm initially downplayed the issues as minor.
Contractual scrutiny and governance backlash
Public sector experts and government officials have strongly criticized the flaws. Professor Stephen Bartos from the University of Canberra noted that the extensive errors discredit the study, suggesting that a completely new report or a ground-up manual rework without AI is required. State opposition members labeled the document "AI slop," while WA Tertiary Education Minister Tony Buti initially characterized the mistakes as typos before expressing anger and disappointment.
A spokesperson for the WA Department of Energy and Economic Diversification confirmed that formal steps are underway to recover costs from Deloitte. While state officials contend the errors do not alter the underlying economic modeling or core findings, Deloitte will review and reissue the report. Meanwhile, Murdoch University Vice-Chancellor Professor Andrew Deeks opposed the merger initiative entirely, calling the report "another fail for Deloitte" and arguing that the merger would drain resources while restricting options for students. Professor Christian Downie of the Australian National University urged governments to recalibrate outsourcing contracts with consulting firms, calling for greater transparency regarding how AI is applied and stricter contractual penalties for flawed outputs.
What it means for developers
This high-profile failure illustrates the risks of deploying generative AI tools in high-stakes analytical workflows without rigorous human verification. When AI systems are permitted to generate citations or summarize financial data unchecked, hallucinated metadata and arithmetic errors can easily undermine client trust and compromise multimillion-dollar decisions.
For software engineers building enterprise applications, this incident underscores the necessity of implementing strict validation layers, schema enforcement, and multi-model benchmarking. Developers building AI features can test and compare outputs from leading systems—including Claude, GPT, Gemini, Grok, and DeepSeek—cheaply through a single API at https://apixoai.online. Establishing robust guardrails, automated unit tests for math operations, and retrieval-augmented generation (RAG) validation remains essential to prevent raw model outputs from reaching production reports.
Source: ‘Extremely disappointed’: WA government seeks reimbursement from Deloitte over flawed university report — The Guardian AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
