PwC AI reports reveal risks in consulting's AI quality push

PwC AI reports in the Middle East face scrutiny after GPTZero found fake footnotes and unsupported citations in four publications.

The documents were produced over the past two years by PwC Middle East and covered artificial intelligence, public-sector services, electric vehicles and autonomous vehicles, according to GPTZero researchers. The issue matters because the firm sells advice on AI adoption while also presenting public research meant to support client work.

Citations under review

GPTZero said many citations in the four reports did not support the claims attached to them. Some links led to pages that lacked the stated evidence, while others pointed to web addresses that no longer worked or could not be verified from the material provided.

One cited academic study on Riyadh air quality appeared not to exist, according to the researchers. GPTZero said it found no matching paper in the named journal and no trace of the work under the authors cited in the report.

PwC Middle East said it is reviewing the materials and changing some references. The firm said it "takes the accuracy of our published research seriously and is updating a limited number of supporting citations" in the reports identified by the researchers.

The firm also said, "Consistent with our approach to responsible AI, we have quality control processes for research and content development we expect all our people to adhere to." PwC did not explain in the source material how the faulty references entered the reports.

A loan tool example

One report cited a Medium post by a teenage blogger with 280 followers as support for a JPMorgan automation project described as a "real world success story" for agentic AI. The project involved automated checks of commercial loan agreements and was first reported in 2017, before the public launch of ChatGPT made generative AI a corporate priority.

That timing creates a classification problem for PwC's framing. A pre-ChatGPT automation system can still be relevant to AI adoption, but presenting it as a clean example of newer agentic AI risks blurring older rules-based or machine-learning projects with newer autonomous systems.

GPTZero researcher Paul Esau said inconsistent sourcing patterns were a warning sign. In one report, he said, the claim that human error causes 90% of traffic accidents appeared three times in close proximity, once with a footnote, once without one and once with two different citations.

"This claim isn’t fake, but no human is going to cite the same fact three times in two pages using three different sources,"

Esau said, according to the source material.

Another report referred to PwC's own survey of Middle Eastern chief executives, in which 70% were said to expect generative AI to have a significant effect on their business. The footnote attached to that claim linked instead to a media item that did not mention the survey, according to the researchers.

Consulting controls face test

The findings add pressure on large advisory firms that are encouraging clients to use AI while asking their own staff to adopt the tools internally. GPTZero's earlier reviews led EY and KPMG to withdraw reports that the research group said contained similar AI-related errors.

The immediate effect for PwC is reputational rather than financial, based on the information available. Research papers are not audited financial statements, but they help consultants sell expertise, and weak sourcing can make AI governance advice harder to defend with clients.

If PwC's review is limited to correcting references, the global effect is likely to be tighter disclosure and citation checks around AI-assisted research rather than a broad retreat from generative tools. That path would leave PwC able to keep using AI in content production, while the consulting sector shifts more work into verification before publication.

If clients or regulators treat the errors as evidence of poor AI controls, the consequences could widen. PwC would face harder questions on responsible AI advice, and rival advisory firms would have an incentive to advertise stricter review systems as part of their own AI consulting pitch.

The open questions are practical: how much of the text was generated or edited by AI, who approved the reports, and whether the firm will disclose a clearer review process. For the wider industry, the mechanism is simple: AI can speed research production, but unsupported citations turn speed into liability when clients are paying for judgment.