The employee responsible for botched welfare report that was packed with AI errors will depart the organisation
The unnamed partner in charge of Deloitte’s welfare report – which contained multiple AI-generated mistakes – is exiting the firm after Deloitte said it would impose disciplinary measures over how the report was produced.
Deloitte has apologised, issued a corrected version of the report and returned $98,000 of the $440,000 it was paid.
The Department of Finance is reportedly reviewing the standard of Deloitte’s work across the wider public service and may pursue further repayments if Deloitte is found to have breached the government’s supplier code of conduct.
The botched welfare report for the federal government has become a cautionary tale about AI governance, transparency and staff behaviour – especially relevant for HR leaders setting AI policies.
The firm used generative AI in ways that went beyond what the client had approved, failed to clearly disclose that use, and then misrepresented what it had told the department. The fallout has included financial penalties, reputational damage and tougher expectations from government on AI transparency.
What happened
Deloitte was hired by the Department of Employment and Workplace Relations (DEWR) to review the welfare compliance framework, under a $440,000 contract. When fake academic references and a fabricated legal quote were exposed in the published report, Deloitte initially told the department the issues were confined to citations and footnotes.
Internal correspondence later showed:
- Deloitte told the Finance Department that DEWR had been advised that generative AI caused the errors
- A DEWR deputy secretary contradicted this, stating that while Deloitte apologised for citation errors, it never disclosed that generative AI tools were responsible
- Deloitte subsequently walked back its claim and admitted DEWR had not been told AI was the cause
This sequence highlighted not just technical failures, but a breakdown in communication and disclosure.
Unapproved AI use and “scope creep”
DEWR had explicitly allowed Deloitte to use department-licensed AI tools within DEWR’s secure environment to assist with technical work such as analysing code and workflows.
However, Deloitte staff also used open, external generative AI tools (including ChatGPT and MyAssist) for writing-related tasks like summarising legal cases and formatting citations – activities not covered by the original approval and not clearly disclosed to the client.
Later, Deloitte described its tools more narrowly as a locked-down Azure OpenAI GPT‑4o system operating inside DEWR’s environment, but an internal email acknowledged that the open tools were not subject to the same controls. This gap between sanctioned and actual use is exactly the sort of “AI scope creep” that organisations increasingly face when employees rely on convenient tools under time pressure.
Consequences and criticism
Because of the errors and disclosure failures, DEWR requested almost $100,000 back from Deloitte and confirmed that the corrected report, republished in late September, was checked and cited manually with no AI used.
The department retained the report’s main findings and recommendations but acknowledged earlier footnote and reference problems. Deloitte has insisted it stands by the substance of its work and that the corrections do not affect conclusions.
Critics remain sceptical. Academic Dr Christopher Rudge, who first exposed the fake sources, argued that replacing fabricated references with multiple new ones did not clearly show the original claims were evidence-based. In Parliament, Senator Deborah O’Neill condemned Deloitte’s conduct as a “human intelligence problem”, suggesting buyers should verify who is actually doing the work and whether AI is involved – and wryly noting that a ChatGPT subscription might be cheaper than hiring a big consultancy.
Rising expectations on AI transparency
In response, the federal government has begun tightening rules for AI use in public-sector consulting. Major firms are being told to disclose where generative AI is used in contract work, and agencies are starting to embed specific AI transparency clauses into statements of work.
Procurement specialists argue that traditional contracts are not built for tools that generate content as well as analyse data, and that standards must be upgraded. Internationally, regulators and professional bodies are warning that AI “hallucinations” pose particular risks in fields like HR, law, audit and policy – where clients assume that quotations and sources have been rigorously checked.
Implications for HR
While this saga is playing out between a consultancy and government, the core issues sit in HR’s domain: policy, behaviour, accountability and culture.
Key lessons include:
- AI policies must be specific and operational: DEWR authorised limited AI use on code analysis within a secure environment. Deloitte staff extended AI use to content and citation tasks with less control. HR policies and training need to draw clear lines between approved and prohibited uses, not rely on vague, aspirational guidance.
- Disclosure is a people risk, not a technicality: The central failure was not just AI hallucinations but incomplete and incorrect statements about what had happened. HR teams overseeing ethics, conduct and whistleblowing should treat AI use like any other area where honesty and escalation are critical: staff must understand that concealing or downplaying AI involvement is unacceptable.
- Vendor and partner oversight now includes AI practices: Senator O’Neill’s challenge – to ask who is doing the work and whether AI is used – applies directly to outsourced HR analytics, recruitment technology and culture reviews. HR leaders should be asking external providers to explain their AI use, controls, training and verification processes.
- Training needs to build judgement, not just tool skills: Deloitte has argued that its conclusions remain sound. But the controversy shows that even if outcomes appear intact, poor judgement around how AI is used and checked can inflict serious reputational harm. HR development programs must focus on when not to use AI, when to double-check it, and when to escalate concerns.
Where HR goes from here
Most large organisations are already using generative AI in HR and people functions – drafting policies, summarising legal advice, screening CVs, interpreting survey results. The Deloitte case is unlikely to halt that shift, but it should reshape how it is governed.
For HR leaders, the episode is a prompt to ask blunt questions:
- Exactly where is AI being used in our HR and people processes?
- Which tools are formally approved – and which are being used informally?
- Do our people know what must be disclosed to clients, employees, unions and regulators when AI is involved?
- If we faced a Deloitte-style incident, could we quickly explain who did what, which tools were used, and why?
Ultimately, it was people – not an algorithm – who chose the tools, signed the contract, sent the letters and omitted key details. In an era where HR is expected to own culture, ethics and workforce capability around AI, that human factor may be the most challenging and important lesson of all.
