40 of 45: The KPMG Citation Failure and the Limits of Provenance

On or about June 11, 2026, KPMG International withdrew its October 2025 report Total Experience: Redefining Excellence in the Age of Agentic AI. The next day, AI-detection firm GPTZero published the investigation that triggered it: of the report's 45 citations, only 5 were fully accurate, and 40 of the 45 citation titles were fabricated. The Financial Times independently verified the errors and obtained on-record denials from UBS, NHS Greater Manchester, Swiss Federal Railways, and Transport for London.

A pattern, not an outlier

The KPMG withdrawal sits inside a 2025–2026 wave of fabricated AI-generated content reaching professional and official outputs — episodes involving other Big Four firms, a major law firm, a draft national AI policy, and accepted machine-learning conference papers. It is unfolding precisely as content-provenance instruments mature: C2PA Content Credentials (specification v2.2, May 2025), IETF SCITT, Japan's Originator Profile, and the EU AI Act's Article 50 transparency obligations for generative outputs.

What provenance verifies — and what it doesn't

Here is the uncomfortable line this incident draws, and it applies to VAP as much as to C2PA: provenance frameworks verify origin and integrity, not truth. A cryptographically signed, timestamped, hash-chained record can establish which model generated a passage, which retrieval events fed it, which human approved it, and that none of those records were altered afterward. It cannot establish that a cited source exists. Article 50, likewise, imposes marking obligations on providers of generative systems — not a truthfulness duty on a published PDF.

What the Content/Creative AI Profile (CAP v1.0) addresses is therefore narrower and, we would argue, more honest: after a failure like this one, an anchored drafting pipeline makes it auditable and attributable where the fabrication entered — which generation step produced the citations, which checks ran or did not run, which reviewer signed off. In the actual incident, that attribution work had to be done from the outside, by a third-party detection firm and a newspaper, against records that carried no independent integrity guarantees.

What VAP would not have done

VAP is not a fact-checker. It would not have detected the fabricated citations, and it does not prevent hallucinated content from being generated or published. Conformance would have changed one thing only: the evidentiary quality of the record of who and what produced and approved the text.

Per the normative non-guarantee clause, conformance “…does not warrant the correctness, fairness, or safety of the underlying AI decisions — only the integrity, completeness (at anchor granularity), and attributability of their records. VAP generates evidence; competent authorities and courts evaluate it.”
VAP v1.2 §1.6 (English text is normative)

Verified against: GPTZero's published investigation (2026-06-12); Financial Times reporting with on-record denials from the four named organizations; the withdrawn report's publication metadata (KPMG International, October 2025).