A KPMG Report Praised AI. Then Its AI Faked 40 of 45 Sources.
In October 2025, KPMG published a glossy report on how agentic AI was redefining customer experience. By June 13, 2026, the firm had quietly pulled it. The reason has become its own punchline: a report praising the power of AI turned out to be riddled with AI hallucinations.
An investigation by the AI-detection firm GPTZero found that of the report's 45 citations, only five pointed to real, intact sources. The rest were invented, stitched together, or too vague to verify. It is the latest — and most ironic — entry in a fast-growing list of consulting AI failures, and the clearest case yet of what these incidents actually have in common.
What actually happened
The report, Total Experience: Redefining Excellence in the Age of Agentic AI, framed agentic AI as a driver of better customer and operating outcomes. The trouble started when several organizations recognized themselves in it and pushed back. UBS told the Financial Times the information about it was "factually incorrect." The NHS and Transport for London disputed how their work was portrayed.
Then GPTZero published its citation audit. The numbers are striking: of 45 citations, 40 had fabricated titles. Another 28 paraphrased real titles or bolted invented components onto genuine sources, and 12 were phrased too vaguely to verify at all. Roughly half of the factual claims those citations were meant to support turned out to be false or misattributed.
GPTZero gave the behavior a name: "vibe citing" — the citation cousin of vibe coding, where a model stitches fragments of real sources together and invents the rest. KPMG removed the report and said it was reviewing the circumstances of its publication.
KPMG is not alone
If this were a one-off, it would be embarrassing but forgettable. It isn't.
Deloitte Australia delivered a 237-page government report containing up to 20 errors, including fabricated academic papers and a made-up extract from a court case. The firm refunded part of its AU$439,000 fee.
EY withdrew a report after a probe found hallucinated citations, including a reference to a McKinsey study that did not exist, and later saw a separate cybersecurity report pulled for the same reason.
A top law firm apologized to a New York court after an AI-assisted filing misquoted the U.S. Bankruptcy Code.
Different firms, different documents, same failure. That pattern is the real story.
The real lesson: this is a grounding problem, not an "AI problem"
It is tempting to read these incidents as proof that AI is unreliable. That misses the point. Large language models are built to produce fluent, plausible text. Left to generate on their own, they will happily invent a citation that looks exactly right — correct author, plausible journal, convincing year — because nothing in the process forces the output to connect back to a real, retrievable source.
That missing connection is the whole problem. "Vibe citing" is simply what happens when generation is decoupled from evidence. The fix is not a smarter model. It is an architecture that refuses to output anything it cannot trace to a verified source.
What source-grounded document AI looks like
In document workflows, grounding comes down to three design choices:
Extract from the actual source document, and tie every output back to its exact location in that document — so a human (or an auditor) can click from a result straight to the page it came from.
Run a verification step before anything is published, instead of trusting a single generation pass.
Keep the document in a controlled environment, so sensitive material never has to leave your network to be processed.
This is the design principle behind source-grounded document AI. Korea Deep Learning's DEEP Agent, for instance, ties every extracted field back to its position in the source document and runs a verification pass before output — and it runs on-premise, so sensitive files never leave the internal network. The point is not the product; it is the principle. An audit trail that links every answer to its source is what separates a system you can defend from one that quietly vibe-cites.
Why grounding is now a procurement question
For any enterprise buying or deploying AI, the KPMG episode turns a technical nicety into a purchasing criterion. The question to ask a vendor is no longer "how accurate is your model?" but "when your system gives me an answer, can you show me exactly where it came from?"
Systems built this way already exist and are measurable. Korea Deep Learning's vision-language OCR ranked first in the English category of OCRBench v2 (68.1) — ahead of Google Gemini and GPT-4o — at 99% accuracy, and is deployed across 80+ public and financial organizations where a wrong, ungrounded answer is not an option. Grounding, in other words, is not a constraint on performance. It is what makes performance trustworthy.
The firms caught out by AI hallucinations did not lack intelligence in their tools. They lacked evidence. The lesson for everyone else is to build — and buy — for evidence first.
Call to action
Evaluating document AI? Start by asking one question: how does every output trace back to its source?
For more, see how poor document parsing creates downstream hallucinations and the real lesson from the Deloitte and EY cases
Sources
The Register (2026-06-12). KPMG's AI report becomes an accidental demo of AI hallucinations. https://www.theregister.com/ai-and-ml/2026/06/12/kpmgs-ai-report-turns-into-a-demo-of-ai-hallucinations/5255029
GPTZero. Chasing the Hallucinations: KPMG's "Redefining Excellence". https://gptzero.me/news/investigations-kpmg/
Technology.org (2026-06-16). KPMG Pulls Agentic AI Report Riddled With Hallucinations. https://www.technology.org/2026/06/16/kpmg-pulls-agentic-ai-report-hallucinations/
Computing (2026). EY cybersecurity report pulled after probe finds 'AI hallucinations'. https://www.computing.co.uk/news/2026/ai/ey-cybersecurity-report-withdrawn-ai-hallucinations
Fortune (2025-10-07). Deloitte to refund the Australian government after AI hallucinations in a $440k report. https://fortune.com/2025/10/07/deloitte-ai-australia-government-report-hallucinations-technology-290000-refund/