Another Big Four firm caught publishing reports with sources that do not exist
GPTZero found fabricated citations in four PwC Middle East reports. One is 84 percent likely to be entirely AI-generated. KPMG, Deloitte and EY got there first.
After KPMG, it is PwC’s turn. The detection company GPTZero says it found fabricated sources and false claims in four PwC Middle East reports published between 2024 and 2026. The worst of them, a report titled “Transforming Governance,” scores as 84 percent likely to be entirely AI-generated by GPTZero’s own tool. PwC uses that report to promote a product called “Citizen Pulse,” claiming governments in Denmark, Saudi Arabia, the United States and Australia rely on it. GPTZero found no evidence supporting those claims.
The pattern GPTZero describes is worth a name, and they gave it one: “vibe citing.” References are attached loosely, often with incorrect titles, URLs or authors, and frequently do not actually support the claim they sit next to. In the reports examined, the documents with the highest AI-generation scores also had the most false sources. That correlation is the interesting part. It suggests the citations were not checked because nobody in the chain treated them as claims that needed checking.
PwC Middle East told the Financial Times it takes accuracy seriously and is “updating a limited number of supporting citations.” The firm did not explain how the errors got in. GPTZero has run the same exercise on KPMG, Deloitte and Ernst & Young, and found fabricated sources at all three. Earlier investigations led EY and KPMG to retract reports.
Whether this is really an AI problem is a fair question. Consultancies were producing thinly sourced thought-leadership material long before language models existed, and a fake citation is a fake citation regardless of who typed it. What AI changed is the cost. Producing a hundred pages of confident, well-formatted, plausibly-referenced prose used to take a team and a budget, which imposed a natural limit and usually a review step. Now it takes an afternoon, and the review step is the thing that quietly got dropped. Language models generate citations the same way they generate everything else, by predicting what a plausible reference looks like, not by looking one up. Unless the tool is explicitly connected to a search index and the output is verified, plausible is all you get.
What this means for you: Two practical takeaways. First, if you read industry reports as evidence for anything, spot-check two or three citations before you quote the conclusions. It takes five minutes and the failure rate is apparently not small. Second, if you use AI to draft anything with references, treat every citation as unverified until you have opened it yourself. This is the single most reliable way current models embarrass their users, and it is entirely preventable. The rule is simple: the model can find the argument, you check the sources.
Sources
Computing's biggest professional body is debating whether to open its library to AI
ACM leadership argues that keeping the Digital Library out of reach of language models would leave computer science research invisible in an increasingly AI-mediated world. No decision has been made.
