Deloitte forced to refund Aussie government after admitting it used AI to produce error-strewn report (www.techradar.com)

Deloitte admitted it used generative AI (now disclosed as GPT‑4o) to produce a government report for Australia’s Department of Employment and Workplace Relations, but failed to apply adequate safeguards. The draft contained fabricated citations, false footnotes and a made‑up court quote — errors flagged by researchers — so Deloitte agreed to refund the final installment of a AU$440,000 contract while the department corrected the report, removing more than a dozen bogus references, fixing typos and rewriting sections. The updated document now discloses the model use and says substantive findings remain unchanged, but the episode forced an embarrassing retraction and public apology. The incident is a cautionary case for the AI/ML community about the real‑world costs of hallucinations and poor governance. Technically, it underlines the limits of closed‑loop LLM outputs (even from top models) without retrieval augmentation, citation verification, human expert review, provenance tracking and red‑teaming. For consulting firms and public‑sector procurement this raises legal, reputational and compliance risks: documenting model choice and prompts, using grounded sources or RAG pipelines, automated citation checks, and audit trails should be standard. The affair reinforces calls for transparency and operational guardrails when deploying generative systems in high‑stakes policy work.