PwC is now the fourth Big Four firm caught publishing AI generated hallucinations in its own reports

[Rory93]

The Financial Times has verified an investigation by AI detection outfit GPTZero finding that four PwC Middle East thought leadership reports, covering agentic AI for corporates, public service guidance for governments, and electric and autonomous vehicle forecasts, are riddled with fake footnotes, misattributed claims and citations pointing to sources that simply do not exist

Some of the specific examples are almost comically bad, one report cites a teenage blogger with 280 followers on Medium as its source for a well known JPMorgan initiative that was actually first reported back in 2017, five years before ChatGPT even existed, and a supposed academic paper on air quality in Riyadh appears to have been invented entirely, there is no trace of it in the journal it claims to be published in or from any of the credited authors

A GPTZero researcher pointed to a more subtle but arguably more damning tell, one claim about human error causing 90 percent of traffic accidents shows up three times across two pages of the same report, first with a footnote, then with none, then with two footnotes citing two different sources, which is exactly the kind of inconsistency no human writer working carefully would ever produce but that falls right out of stitching together AI generated drafts without close review

PwC is not an isolated case here either, GPTZero's earlier work already forced EY and KPMG to retract similar reports over the past year, and Deloitte had to issue a partial refund to the Australian federal government last October after one of its reports was found to contain AI driven errors, so this is now a pattern across essentially the entire Big Four rather than a one off embarrassment

The irony is what makes this land so hard, these are the exact firms charging premium rates specifically to advise other companies on how to adopt AI responsibly, PwC's own US chief executive was telling Business Insider the same week that AI will just expose how broken your processes already are, and his own firm's citation pipeline turned out to prove that point better than he probably intended, PwC told the FT it takes accuracy seriously and is updating a limited number of citations, without explaining how the errors got published in the first place

7E149Ac8 2Ce2 4266 8940 192F9821B33Cwww.ft.com

[Thomas75]

Citing a teenage Medium blogger with 280 followers as evidence for a major JPMorgan initiative is the detail that should be a case study in every AI governance training deck going forward, that is not a subtle failure

[WaveFunction]

The three different footnotes for the same 90 percent stat across two pages is honestly the smoking gun here, no human editor reviewing that draft carefully would ever let that inconsistency ship, that is a tell of zero human review rather than just AI assistance

[Wizard72]

Four Big Four firms caught in roughly a year is not a coincidence anymore, this is what happens industry wide when everyone races to show clients they're AI forward without actually building the verification step that has to sit underneath any of it

[QuantumOracle45]

Consulting firms selling AI implementation advice getting caught not proofreading their own AI output is peak 2026 irony, hard to trust the advice on responsible AI adoption from anyone who can't even catch a citation to a page that doesn't exist

[Will81]

Genuinely curious how many other consulting and research reports across other industries have similar issues sitting undetected simply because nobody has run them through a hallucination detector yet, this feels like the tip of something much bigger

[Luke_67]

The CEO quote about AI just exposing how bad your process already is aged incredibly poorly within the same news cycle, sometimes the universe really does hand you the perfect self own

[Tyler_16]

What worries me more than the reports themselves is GPTZero's point about indexing, once a hallucinated claim like the fake Citizen Pulse framework gets cited by a Big Four firm, other AI tools start repeating it as fact simply because it's now keyword adjacent and machine readable