Purportedly AI-Assisted Citation Errors Allegedly Found in Newfoundland and Labrador's 2025 Health Workforce Report by Deloitte

A Deloitte-authored Health Human Resources Plan released by Newfoundland and Labrador in May 2025 was later reported by The Independent to contain several inaccurate or apparently non-existent research citations. Researchers named in the disputed citations reportedly denied authoring the referenced papers. Deloitte stated that AI was selectively used to support some citation work and that it is revising the report.

Risk classification

• Primary risk domain: 3 Misinformation
• Primary risk subdomain: 3.1 False or misleading information

The AI system generated false and fabricated academic citations, which were then published in official government policy reports, spreading misleading information.

Additional risk subdomains

• 7.3 Lack of capability or robustness: The AI system failed to perform reliably by hallucinating non-existent academic papers and citations.

Causal factors

• Entity: AI
• Intent: Unintentional
• Timing: Post-deployment

The fabricated citations were an unintentional outcome of using the AI system to support research, occurring after the model was deployed.

EU AI Act risk tier

• Risk tier: 3 Limited Risk

Limited Risk: The system used was Azure OpenAI, a generative AI language system used to generate text and citations, which falls under the category of AI-generated content requiring transparency.

AI system and alleged parties

• AI system: Azure OpenAI (Microsoft)
• AI purpose: Writing Assistant; Technical Text Generation
• Behaviour type: Assistant
• Alleged developer: OpenAI
• Alleged deployer: Deloitte
• Alleged harmed parties: Martha MacLeod, Healthcare policymakers, Government of Newfoundland and Labrador, General public of Newfoundland and Labrador, General public of Canada, General public, Gail Tomblin Murphy, Epistemic integrity, Canadian Journal of Respiratory Therapy

Harm severity

Severity ratings are based on this scale.

Highest direct severity in any category: Substantial. Severity is scored from Negligible to Catastrophic in each harm category, for harm the reports describe as caused directly or indirectly by the AI system.

• Physical: direct Negligible, indirect Negligible
• Infrastructure: direct Negligible, indirect Negligible
• Property: direct Negligible, indirect Negligible
• Financial: direct Substantial, indirect Negligible
• Environmental: direct Negligible, indirect Negligible
• Malicious content: direct Negligible, indirect Negligible
• Differential treatment: direct Negligible, indirect Negligible
• Civil rights: direct Negligible, indirect Negligible
• Democracy: direct Negligible, indirect Negligible
• Privacy: direct Negligible, indirect Negligible
• Psychological: direct Negligible, indirect Negligible
• Epistemic: direct Minor, indirect Negligible
• Child sexual exploitation and abuse: direct Negligible, indirect Negligible

Financial

Reported: The report explicitly describes financial losses associated with the cost of the commissioned reports, totaling approximately $1.89 million across two occurrences, averaging $945,000 per occurrence.

Directly caused: The Canadian government paid nearly $1.6 million for the flawed report, and the Australian government paid $290,000 (though a partial refund was later negotiated).

Inferred additional harm: Additional financial losses may include the internal government resources required to review, correct, and republish the reports.

Epistemic

Reported: The report explicitly describes epistemic harm through the generation of fabricated academic citations and non-existent research papers.

Directly caused: The AI system generated multiple false citations, attributing non-existent papers to real researchers and fabricating a federal court judgment quote.

Inferred additional harm: The inclusion of false citations in public policy documents pollutes the broader academic and policy information ecosystem.

People affected

• Occurrences reported: 2
• People reportedly harmed: 2
• People reportedly exposed: 520000

Potential causes

Management

• Weak Internal Governance: Deloitte failed to embed effective guardrails for responsible AI deployment.
• Prioritizing Speed Over Quality: Management failed to ensure rigorous validation of high-cost public reports.

Technology

• Generative AI Hallucinations: Azure OpenAI system generated nonexistent academic papers and fabricated quotes.
• Unreliable Citation Tooling: Selective use of GenAI to support citations introduced fabricated references.

Data Inputs

• Lack of Source Verification: References cited were not cross-checked against actual academic databases.
• Unvalidated Evidence Base: The evidence base informing the policy reports was not properly validated.

Human Factors

• Overreliance on AI Outputs: Consultants accepted AI-generated research citations without manual verification.
• Deficit in AI Literacy: Staff lacked adequate understanding of generative AI hallucination risks.

Process and Methods

• Inadequate Review Processes: Fact-checking processes failed to detect fabricated academic citations.
• Lack of Disclosure Protocols: No process to disclose the use of generative AI in commissioned reports.

Regulatory Environment

• Lack of Government AI Policies: No provincial policies governing the use of AI in third-party reports.
• Insufficient Public Oversight: No regulatory mechanisms to audit AI-assisted government-commissioned work.

Information quality

• Classification confidence: High
• Reason for confidence: The reports provide clear, consistent details about the fabricated citations, the cost of the reports, and Deloitte's admission of selective AI use in both Canada and Australia.
• Ambiguities identified: The exact prompts used and the specific model version of Azure OpenAI are not fully detailed.

Deloitte used Azure OpenAI to support drafting public policy reports for Canadian and Australian government bodies, resulting in multiple fabricated academic citations. While the incidents highlight growing risks of AI hallucinations in public sector decision-making, the national security impact is minor, representing localized financial loss and minor policy-drafting integrity issues rather than a hostile threat.

• Overall national security impact: Minor
• Scope: Multiple nations
• Primary target: Canada
• Other affected: Australia
• Alleged perpetrator: Deloitte

Threat characteristics

• Imminence: Long-term. Reflects an ongoing strategic concern regarding AI reliability in public policy drafting rather than an immediate crisis.
• Autonomy: Human-supervised. The generative AI functioned as a writing assistant, but human consultants failed to properly oversee, verify, and edit the generated outputs before publication.
• Novelty: Evolved capability. Demonstrates an evolution of standard LLM hallucination risks into high-stakes, high-value government consulting and public policy formulation.

Impact by dimension

• Physical security: Negligible. The incident involved fabricated citations in policy reports and posed no threat to physical systems, critical infrastructure, or human safety.
• Information security: Minor. While AI-generated fabrications occurred, this was an unintentional consulting error rather than a hostile disinformation campaign or intelligence compromise.
• Sovereignty: Minor. The use of fabricated AI citations in a million-dollar government health report briefly compromised the quality of policy input but did not disrupt core state authority or constitutional processes.
• Economic security: Minor. The incident resulted in direct financial losses of approximately 1.6 million Canadian dollars and 290,000 Australian dollars for flawed reports, but did not threaten strategic economic stability.
• Societal stability: Minor. The fabrications could slightly erode public trust in government-commissioned policy research, but did not lead to civil unrest, mass surveillance, or human rights violations.