KPMG AI Report Found to Contain 40 Fabricated or Incorrect Citations Out of 45

A KPMG report examining how businesses are deploying AI contained only five accurate citations out of 45 total — the remainder comprising fabricated references, garbled attributions, and incorrect titles consistent with large language model hallucination, according to findings published by research group GPTZero.

The Financial Times first reported on the citation failures, with Finextra noting the paper contained hallucinated claims about AI use at UBS. TechRadar and City A.M. published their own accounts on 12 June 2026, citing GPTZero's investigation directly.

The failure rate — roughly 89 percent of citations inaccurate — is not a minor quality-control lapse. Citations in professional advisory reports serve a specific evidentiary function: they allow clients, regulators, and downstream analysts to verify the underlying data. A fabricated citation does not merely fail to support its claim; it actively misleads anyone who attempts to trace the sourcing chain. For a Big Four firm whose value proposition rests substantially on analytical rigour, the error rate is operationally significant.

GPTZero's investigation identified hallucinations of the kind that are well-documented in current-generation LLMs when used for research synthesis without retrieval-augmentation or citation grounding. The model confidently generates plausible-sounding reference strings — author names, publication dates, journal titles — that do not correspond to real documents. The tell-tale pattern is consistency of form alongside inconsistency of substance: the citations look right typographically while the underlying works simply do not exist.

Worth flagging here: KPMG's response to a prior citation-error episode points in an interesting direction. When citation problems surfaced in an earlier report, the firm attributed the errors to human error rather than AI generation, according to Going Concern (published October 2025). That framing has become harder to sustain as a general posture. Whether the current report was produced with AI assistance or not, the hallucination fingerprint GPTZero describes — fabricated source strings, garbled attributions, incorrect titles appearing at scale — is not a pattern typically produced by human researchers checking their own footnotes.

The broader context here is an industry-wide stress test that is still in progress. Consulting firms and professional service providers have moved quickly to integrate LLM tooling into research, drafting, and synthesis workflows, often faster than their internal quality frameworks have adapted. The incentive structure is understandable: generative AI compresses the time from brief to draft dramatically. The risk is equally legible: LLMs do not know what they do not know, and without hard citation grounding — retrieval-augmented generation tied to verified corpora, or mandatory human verification of every reference — hallucinated footnotes will propagate into client deliverables.

That risk is not hypothetical anymore. It is in a published KPMG report on AI, a document that was presumably reviewed before release.

For practitioners building or governing AI-assisted document workflows, the technical remediation is reasonably well understood: RAG pipelines anchored to licensed, versioned source material, combined with citation-verification steps that programmatically confirm each reference resolves to a real, accessible document. Neither step is exotic. The harder problem is cultural and procedural — ensuring that speed-to-output pressures do not erode verification discipline, particularly when the draft looks polished and the citations are formatted correctly.

The irony that a report specifically about AI adoption was itself apparently compromised by AI-generated errors will not be lost on the sector. It is also, practically speaking, the worst possible context in which such an error can surface: a document intended to shape client thinking about AI governance, undermined by the very failure mode it might have been expected to address.

KPMG had not issued a public correction as of the reporting date. GPTZero's full investigation is published at gptzero.me.