PwC Reports Cited Sources That Don't Appear to Exist

When you buy a consulting report, you are not paying for the sentences. You are paying for the verification work behind them. An investigation published this week suggests that work did not happen in at least four reports carrying the PwC name.

GPTZero, which builds AI text detection software, examined four "thought leadership" reports published by PwC Middle East between 2024 and 2026. It found hallucinated citations, unverifiable claims, and one report flagged as entirely machine-written once the reference section was excluded.

What the investigation turned up

GPTZero put the odds that PwC's 2025 report "Transforming Governance" was fully AI-generated at 84 percent. Strip out the reference list and that figure rises to 100 percent.

The citations are the more serious problem. That same report promotes a PwC framework called "Citizen Pulse" and states it is in use by the governments of Denmark, Saudi Arabia, the United States and Australia. GPTZero found no public evidence that the product was deployed by any of them, or that those deals exist at all.

Other examples follow the same shape. A footnote in "Building a Cyber-Resilient eMobility Ecosystem" points to a URL that cites ChatGPT. Elsewhere, a JPMorgan agentic AI project is presented as a "real world success story"; the supporting source is a Medium post by a teenage blogger with 280 followers.

Why fabricated citations are the predictable failure

There is a technical reason this keeps happening, and it is worth understanding before you blame a careless analyst.

Language models are better at inventing citations than at inventing prose. A reference has a rigid, highly regular structure: author, year, title, publication, link. The model has seen that pattern tens of thousands of times, so it reproduces the format perfectly. What it does not do is retrieve a real record to fill the format with.

The output is a citation indistinguishable in form from a genuine one. It looks more credible than the surrounding text precisely because it is more structured. The tidiest part of the document is the part least likely to get checked.

A pattern, not an isolated slip

A PwC Middle East spokesperson said the firm takes the accuracy of its published research seriously and is updating "a limited number of supporting citations" in the reports. That is not a denial of the underlying finding.

It is also the second Big Four firm to face this criticism in recent months, after similar findings at KPMG. Two cases in the same tier point at structure rather than individual negligence. Audit and tax work runs through hard quality gates because someone signs with legal exposure attached. Thought leadership is marketing. It carries the same logo and a fraction of the scrutiny, and generative tools multiplied its output volume without multiplying the checks behind it.

What this changes for anyone publishing research

The useful read here is not that AI writing is unreliable. It is that AI shifts where the cost sits. Drafting gets cheaper; verification does not get cheaper at all. Organizations that budget for the first saving and not the second end up in exactly this position.

Three things worth doing this week:

• Spot-check citations in reports you receive. Pick three footnotes at random and try to open them. Five minutes protects a decision you may be about to make.
• Never let a model generate your references. Supply the sources yourself and let the model write around them.
• Flag every named customer and case study for separate review. Inventing a client relationship is the most expensive category of error a published document can contain.

Whatever logo sits on the cover, the burden of checking a claim has quietly moved onto the reader. That is a real cost of the current moment, and almost nobody is pricing it in.

Sources: The Decoder, GPTZero Investigation, City AM, The Irish Times