PwC joins Big Four hallucination list as ghost sources spread
Fabricated citations in Big Four reports are entering AI retrieval systems, threatening the credibility of any brand that cites them.
Key takeaways
- All four Big Four firms have now published reports containing AI-generated fabricated or unverified sources.
- One PwC Middle East governance report scored 84% AI-generated and included unverifiable customer references.
- LLMs treat professional services reports as high-credibility sources, meaning ghost citations in those documents can be repeated at scale.
- Organisations in financial services and the multilateral system that cite Big Four reports in their own content inherit embedded hallucinations.
- Citation auditing before publication is now a baseline requirement, not an optional quality step.
GPTZero scored one PwC Middle East governance report at 84 percent AI-generated. The Decoder reports that the same document promoted a PwC product using customer references that could not be verified. Three other PwC Middle East reports contained fabricated sources or false claims. That completes a sweep: KPMG, Deloitte, Ernst & Young, and now PwC have all published documents in which AI-generated text introduced citations that do not exist.
The pattern is not a series of isolated accidents. It is a structural consequence of how large language models produce text. They generate plausible-sounding references because plausibility is what they are trained to maximise, not accuracy. When a professional services firm deploys an LLM to draft a client-facing report and does not run citation verification before publication, fabricated sources go out on firm letterhead. The Big Four collectively audit or advise the majority of Fortune 500 companies and most of the world's multilateral institutions. Their published reports are primary inputs into strategic decisions, procurement processes, and regulatory submissions. Ghost citations in those documents do not stay contained.
The citation chain breaks in both directions
For B2B brands, the risk runs in two directions at once.
The first is reputational contamination by proximity. LLMs used in AI search products, including Perplexity, ChatGPT with web access, and Google's AI Overviews, crawl and cite authoritative third-party documents to support their answers. Professional services reports, UN agency publications, World Bank working papers, and ISO standards sit near the top of the credibility hierarchy these models assign. If a PwC report that ranked highly in a model's training data or retrieval index contains a fabricated citation linking it to, say, a named industrial client or a policy position, that fabrication can be surfaced and repeated by any LLM that treats the report as a trusted source. The original error multiplies.
The second direction is internal. A financial services firm or multilateral that feeds Big Four reports into its own AI-assisted research workflows, which is now standard practice at institutions from development banks to central banks, inherits the hallucinations embedded in those inputs. The organisation's own AI-assisted outputs then carry citations two or three generations removed from anything real.
GPTZero's findings cover four PwC Middle East documents. That sample is small. What it reveals is not unique to PwC Middle East; it reveals a verification gap that exists wherever AI drafting is used without a structured fact-checking step before publication.
What professional authority costs to maintain
The professional services sector built its credibility, and its pricing power, on the assumption that its published analysis had been checked. A McKinsey Global Institute report or a Deloitte industry survey commands attention precisely because readers assume human experts reviewed the claims. That assumption is now visibly wrong in a growing share of cases.