A Big Accounting Firm's AI Report Was Full of Made-Up Sources

A major report by KPMG about how businesses use artificial intelligence contained almost entirely fake citations. Out of 45 sources listed, only five were real. The rest were invented or scrambled — the kind of errors that happen when AI systems generate plausible-sounding references that don't actually exist.

The Financial Times first reported the problem, followed by coverage from Finextra, TechRadar, and City A.M.. The investigation came from a research group called GPTZero.

Why This Matters More Than a Typo

When a consulting firm publishes a report, the citations aren't just decoration. They let clients and other readers check the facts for themselves. A fake source doesn't just fail to prove anything — it actively misleads anyone who trusts the work and tries to verify it.

KPMG is one of the world's largest accounting and consulting firms. Its reputation depends on getting the details right. An 89 percent error rate in citations is a serious problem, not a minor slip.

How AI Creates Fake References

Large language models — the AI systems that power tools like ChatGPT — can sound very confident when they make things up. They generate author names, dates, and journal titles that look completely real. The citations look correct on the surface but point to documents that don't exist.

This particular failure shows a pattern familiar to AI researchers: when these models are asked to find and summarize sources without built-in safeguards, they sometimes invent references from scratch rather than admit they can't find something.

The Bigger Picture

Consulting firms have rushed to use AI tools in their research and writing processes. The appeal is obvious: AI can produce a draft much faster than a human working alone. But many firms haven't yet set up strong quality checks to catch these kinds of errors before reports go to clients.

The technical fix exists and is not complicated. Systems called "retrieval-augmented generation" tie AI writing tools directly to verified sources — real documents that have been checked. Combined with a human step that verifies each citation actually exists, this problem becomes solvable.

The harder part is not the technology. It's making sure that the pressure to work fast doesn't skip over the verification step, especially when a draft looks polished and professional.

The Irony

A report specifically about how companies should safely deploy AI was itself apparently undermined by the very failure that AI systems are known for. It's the worst possible setting for this kind of error — in a document meant to help business leaders make decisions about artificial intelligence.

KPMG did not issue a public correction at the time of reporting. GPTZero's full investigation can be found at gptzero.me.