KPMG Withdraws AI Report After Fabricated Citations and Hallucinated Case Studies Are Exposed

KPMG has pulled a report on AI adoption by businesses after the document was found to contain AI-generated hallucinations, including fabricated case studies and citations that did not match their stated sources. The Financial Times first exposed the errors; UBS subsequently confirmed the findings, according to a LinkedIn analysis by Lynn Räbsamen, CFA.

The report had been written, at least in part, by generative AI — a fact that became impossible to obscure once the citation trail collapsed under scrutiny. An analysis by GPTZero found that only 5 of the report's 45 citations actually matched their referenced sources, per The Register. That is an 89 percent citation failure rate on a document published under the name of one of the world's largest professional services firms.

The mechanics here are familiar to anyone who has spent time stress-testing large language models. Hallucination in citation-heavy tasks is a well-documented failure mode: the model produces plausible-sounding references — correct author names, credible journal titles, realistic page numbers — that simply do not exist or do not say what the model claims. The problem is compounded when the output is not subjected to line-by-line fact verification before publication. In a research report built around cited evidence, that verification step is not optional; it is the entire epistemic foundation of the document.

What makes this incident notable is the institutional weight behind it. KPMG is not a startup experimenting with AI-assisted content. It is an organization that advises governments, regulators, and major corporations — often on questions involving risk and compliance. A report on AI's business benefits, written substantially by AI, and riddled with fabricated supporting evidence, is a specific kind of credibility problem: the subject matter and the failure mode are identical.

The broader context here is a growing pattern of AI-assisted research outputs reaching publication without adequate human review. This is not unique to consulting firms. Academic pre-prints, market research, and financial analysis have all seen similar incidents over the past two years. What varies is the reputational surface area of the organization involved. KPMG's withdrawal will land differently in boardrooms than a retracted pre-print from an unknown institution.

Worth flagging: the GPTZero citation-match figure — 5 out of 45 — deserves some methodological caution. GPTZero's primary tooling is built around AI-content detection, not citation verification, and the specific methodology behind that count has not been independently detailed in available reporting. The number is striking, but readers should treat it as an indicator of serious problems rather than a precisely audited finding.

That caveat aside, the withdrawal itself is the hard fact. KPMG did not issue a correction or an addendum; it pulled the report. That is a significant response from a firm whose publications carry substantial institutional authority, and it suggests internal review confirmed the problems were not minor or isolated.

For technology professionals advising organizations on AI deployment, the practical signal here is straightforward. Generative AI degrades predictably on citation-dependent tasks when not paired with grounded retrieval — retrieval-augmented generation, or structured verification against a curated document corpus. Using a base LLM to produce a research report and then publishing it without citation-level human review is not a workflow problem; it is a process design failure. The output looks authoritative. The hallucinations are grammatically fluent, stylistically consistent, and contextually plausible. That is precisely what makes them dangerous in a professional document.

The good news, if there is any, is that the detection and withdrawal happened. The FT investigation and the subsequent UBS confirmation are the kind of external accountability that corrects the record before the fabricated case studies embed themselves in secondary citations and slide decks. The correction mechanism worked. The prevention mechanism did not.

KPMG has not, as of reporting, detailed what internal review process the report went through before publication, or what changes it will make to its AI-assisted research workflows going forward. Those answers matter more than the withdrawal itself.