PwC Middle East Reports Contain Fabricated AI Citations
PwC Middle East published four reports between 2024 and 2026 that GPTZero's investigation found contained fabricated citations, unsupported claims, and broken or mismatched references. The Financial Times independently verified the investigation, according to the Australian Financial Review. The reports covered subjects including agentic AI, public services, cybersecurity, and electric vehicles, raising evidence-governance concerns for organizations publishing AI-assisted research.
PwC Middle East published four "thought leadership" reports between 2024 and 2026 that contained fabricated citations, unsupported claims, and references that did not substantiate the accompanying text, according to an investigation by GPTZero. The Financial Times independently verified the investigation, the Australian Financial Review reported.
The reports addressed agentic AI, public services, cybersecurity, and electric vehicles. GPTZero characterized the material as showing a pattern of AI-generated drafting, citing its Hallucination Check and AI Detector results alongside manual review of citations and source claims. An AI-detection score alone does not establish that text was generated by a model, but the investigation's central findings concern verifiable citation failures and claims unsupported by linked sources.
Unsupported claims and unverifiable references
GPTZero identified the 2025 report "Transforming Governance" as the most serious example. The report promoted a purported PwC framework called "Citizen Pulse" and claimed that governments in Denmark, Saudi Arabia, the United States, and Australia used it to improve public services. GPTZero reported that the cited sources did not support those claims and that it found little public evidence of the framework outside the report.
TechSpot reported further examples across the documents:
- •A report cited an alleged internal PwC survey claiming nearly 70% of regional CEOs expected generative AI to redefine their business landscape, while the referenced article did not contain that survey.
- •A cybersecurity-report footnote included a URL carrying the utm_source=chatgpt.com tracking parameter.
- •An AI case study about JPMorgan's commercial-loan review automation relied on a Medium post by a teenager as its sole cited source, according to TechSpot. The underlying JPMorgan project dated to 2017.
- •An electric-vehicle report referenced an untraceable Riyadh air-quality study and made repeated claims that human error causes 90% of road accidents, using inconsistent or absent attribution.
The Australian Financial Review described the reports as materials intended to support consulting business in the Middle East. Its account, drawing on the GPTZero investigation and FT verification, reported fake footnotes, misattributed claims, and unverifiable information.
What the incident means for AI-assisted research
For data and AI teams, the issue is less whether generative AI touched a draft than whether a publication workflow can trace every external assertion to a valid, relevant source. Citation validation needs to test that a URL resolves, that the cited source exists, and that it actually supports the nearby claim. Those checks are distinct from plagiarism screening and AI-text detection.
Organizations using language models for research production commonly face a specific failure mode: fluent prose can conceal invented sources, obsolete evidence, or citations that are real but semantically unrelated. A review process that combines retrieval provenance, claim-level source checks, domain-expert review, and versioned approval records can catch more of those errors than a final copy edit alone.
The reported errors also illustrate a reputational risk for published research in high-trust professional settings. Claims about government deployments, client projects, executive surveys, or performance outcomes require especially strong validation because they can be independently checked and may affect customer, partner, and public-sector relationships.
Key Points
- 1GPTZero identified fabricated and mismatched citations across four PwC Middle East reports, and the Financial Times independently verified the investigation.
- 2The reported errors included unsupported government-use claims, untraceable studies, broken references, and a ChatGPT tracking tag embedded in a citation.
- 3Organizations publishing AI-assisted research commonly need claim-level provenance checks because fluent model output can obscure invalid or irrelevant evidence.
Scoring Rationale
The reporting documents a prominent professional-services firm's publication of research with allegedly fabricated and unsupported citations, a material governance failure for AI-assisted knowledge workflows. It is directly relevant to practitioners building retrieval, research, compliance, and content-review pipelines, though it is not a new model, platform, or regulatory action.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
