Cases › made things up › LR-2026-0316
Deloitte report with AI-fabricated citations for an Australian federal welfare review (DEWR)
Preliminary Why: 1 report: HRD (Human Resources Director) · subject to change · 1 independent report
Status: fix reported
No source states: loss or records, agent tool, could do, deployer, industry.
Severity moderate: tier: the data of others was touched with a bounded scope stated (over a dozen fabricated points attributed to colleagues in one report); no regulator or count of two or more affected parties is stated.
Deloitte produced a report for Australia’s Department of Employment and Workplace Relations (DEWR) on a welfare compliance framework, and used generative AI tools in publishing-related tasks. The report contained fabricated academic sources and a fabricated quote from a leading case. University of Sydney academic Chris Rudge identified the fabricated points. Deloitte first told the department the errors were limited to footnotes and the reference list, and later corrected its statements about whether it had told DEWR that generative AI caused the errors. DEWR requested repayment of $97,587.11 and the final instalment of the contract was withheld. A republished report dated 26 September, released publicly on 3 October, was said to have been reviewed without AI, with citations done manually.
What happened
What was done
“Republished report with review and citations done manually without AI”
Risks
Sorted by rule from the case's fields; not yet read from the sources.
Controls
ApplicationNot effective User policy and training
Control failure chain
L1 · ModelChatGPT and MyAssistModel provider
No source mentions a control at this layer.
Model version not linked to a model card.
L2 · HarnessHarness not namedAgent framework vendor
No source mentions a control at this layer.
L3 · ApplicationDeployer not namedDeployer
Other controls for this risk at this layer, not mentioned in the sources: Grounding and fact checks, Human approval for high-impact actions, Output filtering and redaction, People told they are dealing with AI.
L4 · Infrastructure / targetOrganisation hit
No source mentions a control at this layer.
EffectiveNot effectiveNot effective (bypassed)Did not existNot stated in sourcesNot yet checkedAttacker's ownOrder: model · harness · application · target
Controls for this risk: result in this incident
| Control | Part of the system | Result |
|---|---|---|
| User policy and training | L3 Application | Not effective |
| Human approval for high-impact actions | L3 Application | Not stated in sources |
| People told they are dealing with AI | L3 Application | Not stated in sources |
| Output filtering and redaction | L3 Application | Not stated in sources |
| Grounding and fact checks | L3 Application | Not stated in sources |
Similar incidents
Evidence
No source states: what the AI was doing, when it happened, when it was found, how many records were exposed, the direct loss, a count of the other parties affected, any court case, the cause, what set it off.
Show the full account: 3 items from the primary sources
What happened, in order
Not dated in the sources:
- Deloitte's assurance was later undercut by discovery of a fabricated quote from a leading case on robo-debt.
- A DEWR deputy secretary contradicted Deloitte's account, stating the department was never advised that generative AI caused the errors.
- DEWR requested repayment of $97,587.11.
Show every field, with its source and grade
The AI system
What happened
Impact
Context
Notes on gaps
The year 2025 for the 22 August disclosure is inferred from the article date. Controls were not reported as held because no source states that. The sources do not say whether a fact-check step existed.
LR-2026-0316 · version 1 · published 2026-11-01 after automatic checks; not yet checked field by field by a person · JSON