One firm published a cost survey. Another published reports its own AI made up.

The same week EY put numbers to the AI reckoning, an outside investigation found four PwC reports built on fabricated citations and a framework no government could be shown to have used.

The week set two of the Big Four side by side on the same subject, trust, and let them answer it in opposite ways. EY spent the week on offense, publishing hard survey data on AI cost and arguing that companies must move from adoption to accountability. PwC spent it on defense. An investigation by the AI-detection firm GPTZero found that four PwC Middle East thought-leadership reports published between 2024 and 2026 were riddled with fabricated citations, unverifiable statistics, and misattributed sources.

The details are not subtle. One report, on governance, was assessed as eighty-four percent likely to be AI-generated, a figure that rose to one hundred percent once its reference section was removed from the analysis. It introduced a PwC framework called Citizen Pulse and claimed the product had been implemented by the governments of Denmark, Saudi Arabia, the United States, and Australia. The investigators could find no public evidence that the product was used or that any of those deals existed. A separate cybersecurity report cited a link that pointed to a chatbot rather than to any primary source. PwC told the Financial Times that it takes the accuracy of its published research seriously and is updating a limited number of supporting citations.

It would be easy to treat this as an embarrassment and move on, but the lesson is structural, and it is the same lesson the cost story teaches from the other side. PwC sells assurance. Its entire proposition in this market is that it can stand behind a claim. And it could not keep fabricated citations out of a document with its own name on the cover, in the one category, thought leadership, where the firm controls every input. If the assurance cannot survive its own reports, what exactly is being assured in a client's?

This is the difference between a promise and a proof. A thought-leadership report is a narration: it asserts, and it invites you to trust the letterhead. A proof is a different object, a record of where each claim came from that a reader can follow back to the source, kept in a form that does not depend on the author's good name. The hallucinated citation is not a slip. It is what happens when a narration is generated faster than it can be grounded, and no structural check sits between the draft and the reader.

Consulting-as-Code is built on the opposite default. Every figure traces to the source and the formula that produced it, the record is structured so a reviewer can inspect it, and a claim that cannot be grounded does not survive the process that assembles it. The point is not that PwC used AI. Everyone uses AI. The point is that assurance you can inspect and assurance you are asked to take on faith are not the same product, and this week the difference showed up in public, in the work of a firm that sells the second and called it the first.