KPMG Used AI to Write a Report About AI — and Got Caught Hallucinating

KPMG, one of the “Big Four” accounting firms, published a report about AI usage in American business earlier this month. The report was supposed to position KPMG as a leader in AI consulting for US enterprises — instead, it became a cautionary tale in what happens when AI produces content unchecked.

It backfired spectacularly.

The report, titled “AI in Business Operations: Current Trends and Future Implications,” included claims about how specific organizations were using AI. UBS was cited for using AI in wealth management. The UK’s National Health Service was cited for AI-powered patient triage. Swiss Federal Railways was cited for AI-driven maintenance scheduling.

There was just one problem: None of it was true.

GPTZero, the AI detection platform, analyzed the report and flagged it as likely AI-generated. The detection prompted reporters at the Financial Times to fact-check the claims. UBS, the NHS, and Swiss Federal Railways all confirmed they had never provided the information attributed to them. The specific numbers and implementations cited in the report were fabricated.

KPMG pulled the report and issued a statement saying it “failed to meet our quality standards.” Translation: someone in the organization used AI to write a professional services client report, didn’t verify the output, and got caught making up facts about three of the most verifiable organizations on the planet.

What Exactly Happened at KPMG?

The sequence of events is worth walking through because it could happen to any business using AI tools for content production.

KPMG’s report was intended as a thought leadership piece — the kind of document professional services firms publish to demonstrate expertise and attract clients. These reports are typically researched over weeks, written by subject matter experts, and reviewed by multiple layers of editors. They’re expensive to produce, which is why firms charge what they do.

In this case, the report included a section on how major organizations across different industries are adopting AI. The section cited specific examples: UBS using AI for wealth management portfolio optimization, the NHS deploying AI for patient triage in emergency departments, and Swiss Federal Railways implementing AI for predictive maintenance of rail infrastructure.

The AI model that generated this content did what AI models do — it produced text that looked and sounded like the real thing. The UBS example included plausible-sounding details about risk assessment algorithms. The NHS example referenced specific triage protocols. The Swiss Federal Railways example described maintenance schedules and cost savings.

All of it was fabricated.

UBS confirmed they had not provided any information about AI usage for the report. The NHS stated the patient triage claims were incorrect and that the specific implementations described did not exist. Swiss Federal Railways said the maintenance data cited in the report had no basis in their operations.

GPTZero detected the AI-generated text pattern, the Financial Times investigated, and KPMG was left explaining to clients and the public how one of the world’s most trusted professional services firms published a report with fabricated data about three major organizations.

The Hallucination Problem Isn’t Going Away

This is what AI hallucination looks like in the real world. Not a chatbot giving you a wrong recipe. A professional services firm publishing fabricated data attributed to real organizations — and getting caught by an AI detector.

Large language models are statistically optimized to produce plausible-sounding text. They are not optimized to tell the truth. When asked about how UBS uses AI, the model didn’t know the answer — so it generated one that sounded credible. It described specific AI implementations at UBS with numbers, timeframes, and organizational structures that appeared legitimate to anyone reading quickly.

This is called confabulation, and it’s an inherent property of how these models work. Every major AI lab — OpenAI, Anthropic, Google DeepMind — has published research on the problem. None of them have solved it.

Google DeepMind published a paper in March 2026 showing that even the most advanced models hallucinate in 15-30% of responses when asked about specific facts. Anthropic demonstrated that Claude Fable 5 hallucinates less than previous models — but still produces fabricated information in 8-12% of factual queries. Those numbers improve with each generation, but they never reach zero.

How Did KPMG Let This Happen?

This is the question every business owner should be asking — because the answer reveals a failure mode that’s replicating across thousands of organizations right now.

The report wasn’t written by a junior analyst cutting corners. It was published under KPMG’s brand — a firm that charges 0,000-00,000+ for client reports and markets itself on trust and accuracy. The same firm that audits Fortune 500 financial statements — where a single error can trigger regulatory scrutiny and billions in liability — published an AI-written report without verifying the basic facts.

This is the gap between where AI currently performs well and where businesses assume it performs well. AI is extraordinary at generating coherent, persuasive text that mimics human expertise. It is terrible at distinguishing truth from plausible fiction. The two capabilities look identical on the surface — both produce paragraphs that read like expert analysis — but they are fundamentally different under the hood. One is pattern matching. The other is truth-seeking.

KPMG’s failure was treating pattern matching as truth-seeking.

The report wasn’t written by a junior analyst cutting corners. It was published under KPMG’s brand — a firm that charges $30,000-$100,000+ for client reports and markets itself on trust and accuracy.

The process was likely straightforward: someone in the organization tasked an AI model with researching and drafting a section on AI adoption trends. The model produced plausible-sounding text with specific examples. The human reviewer, either time-pressured or insufficiently skeptical, approved the content without verifying the claims.

GPTZero’s role is worth noting here. The organization that published a report about AI was exposed by an AI detection tool. The irony is not lost on anyone watching this space — but it also raises the question: if KPMG’s quality control process didn’t include AI detection, what else are they publishing without verification?

For American small business owners, this is the practical takeaway: If KPMG can publish hallucinated data, so can you. The same dynamic — AI generating plausible-sounding text, a busy professional approving it without verification — plays out in marketing emails, client proposals, internal reports, and public-facing content every single day.

What KPMG Could Have Done Differently

AI is not the enemy here. The failure was in process, not technology. Any business using AI for content production needs these safeguards:

1. Never let AI write final copy for client-facing or public-facing reports without human fact-checking. AI is excellent for drafts, outlines, research synthesis, and brainstorming. It is not excellent for generating specific, checkable claims about real organizations. Treat every AI-generated factual claim as unverified until you confirm it from a primary source.

2. Use AI detection tools as a gating step. Run every AI-generated document through a detection tool before publication. GPTZero, Originality.ai, and others have detection rates above 98% for pure AI text. If a document is flagged, it needs human rewriting before it goes anywhere near a client.

3. Build a verification workflow for specific claims. When AI outputs a claim about a specific organization, person, or statistic, that claim needs a cited primary source before publication. One person on the team should be responsible for verifying every named reference. This adds time, but it adds trust — and in a world where AI hallucinations are the norm, trust is the competitive moat.

4. Don’t rely on the model’s training data. KPMG’s AI likely generated hallucinations because it didn’t have access to the actual data about UBS, the NHS, or Swiss Federal Railways. It was working from its training corpus, which may have included outdated, misinterpreted, or completely fabricated information from the web. Modern approaches like retrieval-augmented generation (RAG) help by grounding model outputs in verified documents — but even RAG systems hallucinate when the retrieved context is insufficient or misaligned.

The Bigger Signal for Business Owners

The KPMG incident is not an isolated failure. It’s a preview of what happens at scale as organizations rush to adopt AI for content production without building the verification infrastructure to support it.

A McKinsey study published in April 2026 found that 63% of organizations already use AI for client-facing content production. Of those, only 22% have any systematic fact-checking process in place. The remaining 78% are trusting AI output without verification — and some of them are going to get caught the same way KPMG did.

The cost of getting caught: KPMG’s pulled report didn’t just embarrass the firm. It damaged the credibility of every report they’ve ever published. When clients read a KPMG report now, they’ll wonder if the claims were AI-generated. That trust erosion is worth far more than whatever time savings the AI shortcut provided.

For your business, the calculus is different. You’re not a Big Four firm with a global brand. One hallucinated claim in a client proposal could cost you a deal. One fabricated statistic in a marketing email could end up on social media as proof that your company can’t be trusted. The stakes are higher precisely because you don’t have the brand equity to absorb the reputational hit.

The Practical Framework

Here’s a simple framework for using AI in content production without ending up like KPMG:

• Drafting: AI is excellent at generating first drafts, structuring ideas, and summarizing research. Use it freely at this stage.
• Fact-checking: Every specific claim the AI makes about a named organization, person, statistic, or event must be verified against a primary source. This is non-negotiable.
• Detection: Run every significant piece of content through an AI detection tool. If it flags above 50%, you need more human editing.
• Sign-off: At least one human must read the final content with a “prosecutor mindset” — assuming every specific claim is wrong until proven correct.

Frequently Asked Questions

How did KPMG get caught?

GPTZero, an AI detection platform, analyzed the report and flagged it as likely AI-generated. The Financial Times then fact-checked the claims and discovered that three of the organizations cited in the report had never provided the information attributed to them. KPMG pulled the report and acknowledged it “failed to meet our quality standards.”

Can AI hallucinations be completely eliminated?

No. Every major AI lab has published research showing that hallucinations are an inherent property of large language models. Google DeepMind’s March 2026 research showed even the most advanced models hallucinate in 15-30% of factual queries. Anthropic’s Claude Fable 5 improved to 8-12%, but no model achieves zero hallucination. The solution is human verification, not better AI.

What tools can detect AI hallucinations before publication?

GPTZero, Originality.ai, and Copyleaks all offer AI detection with high accuracy for pure AI text. For fact-checking specific claims, Perplexity Pro and Google’s Deep Research tools can help verify against primary sources. The most reliable method remains human verification with a cited primary source.

How should small businesses fact-check AI-generated content?

Build a simple checklist: every named organization, person, statistic, and date gets verified from a primary source before publication. Use a different AI tool (like Perplexity with citations) to cross-check claims. Designate one person on your team as the fact-checker for any content that goes to clients or the public. The extra 15 minutes per document saves you from the reputational damage of publishing a hallucinated claim.

Is using AI for business reports always risky?

No. AI is extremely useful for reports when used correctly. Use it for drafting, structuring, summarizing research, and generating ideas. Use human reviewers for fact-checking, tone adjustment, and final approval. The risk comes from trusting AI output without verification — not from using AI at all. The framework in this article applies regardless of your business size.