Key Takeaways
- Deloitte Australia produced a 237-page report for the Australian government that included fabricated citations and a false court reference created by artificial intelligence (AI).
- The AI used was OpenAI’s GPT-4o model, accessed through Microsoft’s Azure OpenAI Service.
- Deloitte refunded approximately 98,000 AUD from the original 440,000 AUD contract and issued a corrected version of the report.
- The problem was not Azure’s infrastructure but human oversight and failure to verify AI-generated text.
- The case became a global reference for how generative AI should be used responsibly in public-sector consulting.
Story & Details
What Deloitte Did
In early 2025, Deloitte Australia was contracted by the Department of Employment and Workplace Relations (DEWR) to analyze a welfare compliance system and deliver a detailed report valued at 440,000 AUD.
The document contained several fabricated academic citations and even a quote from a court case that never existed. These errors were traced to text produced by GPT-4o, a generative model developed by OpenAI and accessed through Microsoft’s Azure OpenAI Service.
When the inaccuracies were discovered by researchers and media, Deloitte investigated internally and confirmed that parts of the analysis had been generated using AI without sufficient human review. The firm withdrew the document, issued a corrected version, and added a note clarifying that AI had been used in the writing process.
Although the Australian government stated that the main conclusions remained valid, the reputational damage was significant. To resolve the issue, Deloitte voluntarily refunded the final instalment of the contract (around 98,000 AUD) and publicly apologized. Officials later said that future government contracts would require clear disclosure whenever AI tools are involved in report creation.
Azure OpenAI and Its Role
Azure OpenAI Service is Microsoft’s enterprise platform that allows organizations to use OpenAI’s models within the Azure cloud. This setup provides strict data privacy, security compliance, and regional data storage—important for government clients.
The model itself, GPT-4o, is identical to the one used in ChatGPT. The difference is that Azure gives companies private and auditable access through an application programming interface (API) rather than an open public interface. This means that hallucinations—the generation of plausible but false statements—can occur in both environments.
In this case, Azure did not cause the hallucinations; the model’s probabilistic text-generation process did. The incident underlined the need for robust human validation and transparency policies, regardless of how the AI is hosted.
Hallucinations, Responsibility & Oversight
AI “hallucination” occurs when a model produces information that sounds realistic but is factually incorrect. This happens because the system predicts likely word sequences without verifying facts.
The Deloitte report’s false citations and invented legal quote were classic examples of such hallucinations. The real failure, however, was organizational: AI output was treated as authoritative without rigorous review.
Australian lawmakers and experts criticized the oversight, emphasizing that AI tools should support human expertise, not replace it. Deloitte has since revised its internal guidelines, requiring disclosure of generative AI use in all deliverables and mandatory fact-checking before publication.
Despite the controversy, Deloitte continues to expand AI adoption across its operations, citing the event as a valuable lesson in governance and transparency.
Entities & Roles Index
- Deloitte Australia — Consulting firm responsible for producing and correcting the flawed report.
- DEWR (Department of Employment and Workplace Relations) — The Australian government department that commissioned the report.
- Microsoft Azure — Cloud platform providing enterprise access to OpenAI models.
- OpenAI — Developer of GPT-4o, the generative model used in the report.
- Australian Parliament Members — Critics who demanded stricter rules on AI use in public contracts.
- Academic reviewers and journalists — Detected the false citations that triggered the controversy.
Conclusions
The Deloitte–Australia incident demonstrates that even when AI operates in secure enterprise settings like Azure, human accountability cannot be outsourced.
Generative AI remains prone to factual invention, and institutional safeguards must ensure verification and disclosure.
Public trust depends on transparency, and this episode stands as both a warning and a guide for responsible AI adoption in the public sector.
Sources
- https://www.theguardian.com/australia-news/2025/oct/06/deloitte-to-pay-money-back-to-albanese-government-after-using-ai-in-440000-report — describes Deloitte’s partial refund and error disclosures
- https://www.ft.com/content/934cc94b-32c4-497e-9718-d87d6a7835ca — Financial Times analysis of the revised report and contract terms
- https://apnews.com/article/ab54858680ffc4ae6555b31c8fb987f3 — reports on fabricated citations, refund, and correction
- https://www.businessinsider.com/deloitte-australia-issues-refund-ai-assurance-project-2025-10 — summary of the mistake, disclosure, and refund
- https://arstechnica.com/ai/2025/10/deloitte-will-refund-australian-government-for-ai-hallucination-filled-report/ — technical angle on “AI hallucinations” risk
- https://www.youtube.com/watch?v=C17KWJ02Goo — video: “Deloitte delivers report … containing errors”
- https://www.youtube.com/watch?v=vVv2JowMPxw — video: “Deloitte’s AI Mess: $440,000 Report For Australian Govt”
- https://www.youtube.com/watch?v=045GE7zOFhg — video: “Deloitte to refund government after AI mistake”
- https://www.youtube.com/watch?v=UazVan1O4GI — video: “Big 4 new AI rule after Deloitte refund”
My Diary - Leonardo Cardillo 