Why a Major Consulting Firm Just Pulled Its AI Report From the Internet

KPMG, one of the world's largest consulting firms, took down a report about how businesses use artificial intelligence. The reason: the report was full of fake references and made-up case studies that the AI writing tool had created. The Financial Times discovered the problem first, and another major firm, UBS, confirmed it was true.

The report was written mostly by an AI tool called a large language model — the same kind that powers chatbots like ChatGPT. When researchers checked the sources the report cited, they found a shocking problem: only 5 out of 45 references actually existed or said what the report claimed they said. That means 89 percent of the citations were wrong.

This happens because AI language models work by predicting what words should come next, based on patterns in their training data. They are very good at making sentences that sound real and authoritative. But they do not actually check whether what they write is true. If a source does not exist in their training data or if they need to invent a detail, they will make something up — and make it sound convincing. This is called "hallucination" in AI terms.

When you are writing a research report, citations are like the evidence in a court case. Without real, verified sources backing up your claims, the whole report loses meaning. Checking every citation before publishing is not something you can skip. It is the entire point of the document.

What makes this situation different from a student accidentally using a fake source is that KPMG advises governments and big corporations on serious matters like risk and rules. A report about how helpful AI is for business, written by AI itself, and filled with fake evidence, creates a very specific kind of problem. The subject and the failure are the same thing.

This is not an isolated incident. Over the past couple of years, similar problems have shown up in research papers, market studies, and financial reports all over the industry. When a small or unknown organization gets caught, it is usually a quiet correction. When KPMG does it, people in corporate boardrooms take notice.

It is worth noting one thing about the measurement: the tool that found the 89 percent failure rate is primarily designed to detect AI-written text, not to verify sources. The exact method used to check those citations has not been fully explained in reporting. The number is striking enough to show something was very wrong, but people reading this should understand it as a strong indicator of serious problems rather than a precise audit.

But here is the hard fact: KPMG did not just fix the report or add a note. It took the whole thing down. For a firm whose publications influence major business decisions, that is a big step. It suggests the problems inside the company found were not small or scattered — they were extensive.

For people working in technology, the lesson is practical. AI tools tend to struggle with citations unless you use a special technique where the AI searches through real documents before writing — imagine if a person had to consult an encyclopedia before answering every question, rather than just relying on memory. Taking an AI and having it write a research report, then publishing it without someone checking every single source, is not just a bad workflow. It is a broken process. The writing looks official. The fake sources sound authentic. That is precisely the danger.

On a brighter note: the system worked to catch this. Reporters found it, other firms verified it, and the bad report got stopped before the fake examples spread into other presentations and boardroom discussions. The system for catching errors did its job. The system for preventing errors did not.

KPMG has not yet said what checks the report went through before it was published, or how they will change the way they use AI for research in the future. Those answers matter more than the withdrawal itself.