KPMG has withdrawn a report titled Redefining Excellence from its websites after the document was found to contain AI-generated hallucinations, including false citations, fabricated case studies, and claims about AI adoption that named organisations had never made.

The errors were identified by the research group GPTZero and subsequently verified by the Financial Times. Among the hallucinated content were case studies purporting to describe AI deployments by UBS and various health and transit systems. UBS, upon being alerted to the false claims, requested that KPMG remove any reference to the firm. Multiple organisations cited in the report stated that the characterisations of their AI usage were simply untrue.

KPMG pulled the report while it investigates how the publication came to be released, according to The Register. No timeline for that investigation has been made public.

The irony is hard to avoid: a report on artificial intelligence, produced by one of the world's largest professional services firms, became a live demonstration of the failure mode it might otherwise have been expected to address. Hallucination — the tendency of large language models to generate plausible-sounding but factually unsupported output — is among the most discussed risks in enterprise AI adoption. KPMG advises clients on exactly these risks for a living.

That framing, though, is worth holding carefully. The core question here is not whether AI hallucinated; it is whether adequate human review occurred before publication. LLMs produce erroneous output routinely. What separates responsible deployment from negligent deployment is the verification layer that sits between model output and external release. Whatever process KPMG had in place — if any — failed to catch fabricated attributions involving named, verifiable institutions. That is a process failure, not merely a model failure.

The professional stakes are specific. KPMG's core business is trust: clients pay for audit, advisory, and risk services precisely because they expect rigorous factual standards. A published report containing invented case studies attributed to real firms — firms that then had to contact KPMG to demand corrections — is a direct contradiction of that value proposition. The reputational damage is not abstract.

There is also a downstream signal worth watching. Consulting firms and research houses have moved quickly to integrate generative AI into report production, often under pressure to accelerate throughput. The tooling can accelerate drafting, summarisation, and data synthesis meaningfully. But the governance frameworks — editorial sign-off, source verification, legal review of third-party attributions — have not always kept pace with the speed of deployment. This incident will likely prompt a harder look at those frameworks across the sector, not just at KPMG.

For practitioners evaluating AI-assisted research and advisory output, the episode reinforces a basic due-diligence point: named entities, specific statistics, and direct case study attributions in any AI-assisted document should be independently verified before that document is relied upon. The persuasive fluency of LLM-generated prose makes unverified claims harder to spot on a skim read, which is precisely what makes the verification step non-optional.

KPMG has not publicly stated which AI tools were used in producing the report, nor whether the hallucinated content was introduced at the drafting stage or through some other part of the production pipeline. Those details matter for understanding the scope of the failure — and for any firm now auditing its own AI-assisted publication workflows in response.