KPMG pulls an AI-written report about AI after the companies it named said the claims were false

A guideline that everyone agreed with, and no one used

KPMG had a rule. Its spokesperson quoted it, almost wearily: people must use AI responsibly, “including human oversight to validate content and verify independent sources” [7].

The rule was right. The rule was also useless. A report went out the door naming the National Health Service, UBS, Swiss Federal Railways, and Transport for London — and those organisations say the claims about them were false [7]. Nobody at the firm decided to skip the checking. There was no meeting where someone said, “let’s not verify this.” The check just didn’t happen.

That gap — between the rule everyone agreed with and the work no one did — is the thing worth understanding. It isn’t laziness, and it isn’t a bad firm. It’s what happens to a system when the cost of one step collapses and the cost of another doesn’t.

Writing was the bottleneck. Now checking is.

Think about how a consulting report used to get made. Someone gathered facts. Someone drafted prose. Someone footnoted it. That drafting was slow and expensive, so it set the pace of the whole thing. By the time a report was written, it had passed through many hands and many hours — and somewhere in those hours, the facts got read more than once.

The checking wasn’t a separate, scheduled step. It was a side effect of how long the writing took.

Generative AI removed the writing cost. A polished, citation-studded report now appears in minutes [7]. But the model doesn’t gather facts — it predicts plausible words. A hallucination is the model producing confident text that simply isn’t true [7]. The fast part got faster. The slow part — reading every claim against a real source — stayed exactly as slow as it always was.

So the writing no longer drags the checking along with it. The two have come apart. And when a step was never explicitly anyone’s job — when it rode along for free inside another step — it’s the first thing to vanish the moment that ride disappears.

The error that survives a spellcheck

Here is what makes this kind of failure spread. The old mistakes — a typo, a clumsy sentence — looked like mistakes. They were ugly. They caught the eye.

A hallucinated footnote does the opposite. It points to a study with a real-sounding title, a plausible author, a year. It looks more authoritative than a true citation, because the model has learned exactly what authority looks like and reproduces the surface of it perfectly [7]. The output is fluent. Fluency reads as competence. So the one signal a tired reviewer used to rely on — “this looks careful, so it probably is” — now points the wrong way.

This is not unique to KPMG. Last month EY, another of the four biggest firms, withdrew a report with what appeared to be fake footnotes from the same cause [7]. Two of the most careful, most expensive, most reputation-conscious institutions in the business made the same error within weeks. That should tell you the error isn’t about who’s careless. It’s about a tool whose mistakes are camouflaged as competence.

You are downstream of this, even if you never touch the tool

It’s tempting to file this as a story about consultants and move on. But follow where these reports go.

A KPMG report on AI adoption is read by executives deciding what to buy, by boards deciding what to fund, by journalists deciding what’s true, by other firms deciding what’s normal. Some of those decisions touch a hospital’s budget, a transit system’s plans, your bank. The report named the NHS and Transport for London — institutions that serve millions of people who will never read a word of it [7].

And the same pattern is now running everywhere the writing got cheap. The product description, the medical summary, the legal brief, the news article, the school assignment — each one can now be produced faster than it can be checked, by someone who trusted that fluent meant verified. You are not above this system, watching consultants fumble. You’re inside it, reading its output all day, trusting the surface of things because the surface has always been a decent guide.

What seeing the whole actually buys you

The humble move isn’t to stop using the tool, or to trust nothing. It’s narrower and harder: to notice that the check was never really a step, and to make it one — on purpose, by name, with someone’s hours attached.

But notice how little any single seat can see of that. The person who ran the AI saw a finished report. The reviewer saw fluent prose. The executive saw a trusted logo. The reader saw a confident claim. Each of them was looking at something that looked right, and being right is a different thing that no one in the chain was positioned to confirm.

That gap — between looking right and being right — used to be small, because making something look right took almost as long as making it true. The tool just pried those two apart. Knowing they’ve come apart won’t make you certain of anything. It should make you hold what you read a little more loosely — including this.