One of the world's biggest accounting firms published a glossy report about how companies are using AI to serve customers better. Sounds reasonable. Except roughly half of it was made up.
Last October, KPMG dropped a 50-page paper called "Total Experience: Redefining Excellence in the Age of Agentic AI." It looked official. It had 45 citations. It read like the kind of thing you'd see quoted in a dozen other business articles by next week. Then GPTZero, a company that detects AI-generated content, actually checked the citations. Only five of them were real. Twenty-eight pointed to real sources but with fabricated details added. Twelve were too vague to even verify. The rest? Straight hallucinations. GPTZero called this "vibe citing" — basically, the AI just made up references that sounded right.
The fake citations were bad. The fake examples were worse. KPMG claimed Emirates built an AI chatbot called Sara that could rebook passengers' flights. Sara existed, but it was just a basic mobile assistant launched in 2023 with zero ability to change bookings. The company also said UBS had integrated AI agents across its entire investment advisory and compliance operations. UBS told the Financial Times that was "factually incorrect." Swiss Federal Railways supposedly had AI that could optimize passenger trips based on carbon impact and real-time conditions. SBB's response: "not accurate."
Here's the real problem: KPMG is one of the "Big Four" accounting and consulting firms. When firms like KPMG publish research, other companies cite it. Journalists cite it. Analysts build arguments on top of it. Edward Tian, CEO of GPTZero, put it bluntly: error-riddled papers from the Big Four could "poison the well of information" and create what amounts to second-hand hallucinations rippling through business research for years. One bad report from a trusted source doesn't just spread misinformation — it becomes the foundation for other misinformation.
KPMG pulled the report and said it's reviewing what happened. A spokesperson told the Times the firm "takes the accuracy and integrity of its published content seriously." But the damage is already done. The paper was published in October. GPTZero's investigation came out weeks later. By then, who knows how many other reports cited it, how many executives read the fake examples and thought they were learning about real AI deployments, how many business decisions got made based on invented case studies.
The irony is sharp: a report about the benefits of AI got destroyed by AI. And it wasn't even sophisticated hallucinations. These were sloppy ones — made-up citations, misattributed capabilities, examples that didn't exist. If KPMG's AI research tools were doing this, what's happening in the thousands of other corporate reports and analyses being generated the same way right now, with no one checking?