KPMG’s Expert AI Report Was Full of AI-Generated Lies

One of the world's biggest consulting firms published a report on AI. It turned out to be fiction, likely written by an AI. This is the new baseline for corporate incompetence.

A major consulting firm whose entire business is selling expert analysis just got caught publishing chatbot fan fiction. KPMG’s report on “agentic AI” made bold claims about how major companies were deploying the technology. It stated, for instance, that the Swiss bank UBS was using AI agents for investment advice and risk management. The problem? UBS told the Financial Times this was “factually incorrect.” The report wasn't just wrong on a few details; it was fundamentally imagined. Its sources appear to be the statistical phantoms of a large language model. This isn’t an academic debate about the nature of intelligence. It’s a multi-billion-dollar advisory firm, whose name is supposed to be a synonym for trust, passing off demonstrable falsehoods as research.

The failure was a classic hallucination, the most predictable flaw in modern LLMs. These models are generative, not logical; they are built to produce plausible text, not to verify facts. When prompted to write about AI adoption at a firm like UBS, the model assembled a statistically likely sentence by chaining together concepts it had seen in its training data: “UBS,” “AI agents,” “risk management,” “Microsoft.” The resulting prose reads like a corporate press release but has no basis in reality. The error was caught not by KPMG’s internal review but by GPTZero, a company specializing in AI detection, which looks for the unnaturally smooth and predictable statistical patterns of machine-generated text. They flagged it, the FT reported it, and KPMG was forced to pull the report in disgrace. It's the digital equivalent of inventing sources for a term paper, except the machine did the inventing.

The actors here are playing for keeps. KPMG, one of the Big Four firms, just handed its rivals—Deloitte, PwC, EY—a perfect case study in what not to do. Its core product, expert advice backed by rigorous verification, has been fundamentally undermined. For a client paying hundreds of thousands of dollars for a strategy report, the question is now obvious: am I paying for your analysts, or for your intern’s access to a public API? UBS, meanwhile, must now actively fight misinformation about its own technical infrastructure, published by a supposedly trusted partner. The short-term winner is the watchdog, GPTZero, which gets a massive credibility boost. But the real casualty is the already-strained notion of corporate expertise in an age where the pressure to cut costs via AI is overwhelming the processes meant to ensure accuracy.

This incident will not slow the corporate adoption of generative AI. It will, however, create a painful new layer of bureaucratic hygiene. Expect to see armies of junior analysts tasked not with research, but with fact-checking AI-generated drafts. Corporate AI policies will be rewritten with stern chapters on “mandatory human oversight,” a phrase that will become a shield for liability. We will see a boom in “grounded generation” startups that promise to chain LLM outputs to verifiable sources. But those systems will remain more expensive and complex than the cheap, fast, and imaginative models that caused this mess. The temptation to take shortcuts will always be there. So the real question isn't whether the machines will get more truthful. It’s how we verify what they say when the humans in the loop are the ones incentivized to look the other way.