PwC AI Hallucinations: Four Reports and a Big Four Pattern

PwC is the latest — and last — of the Big Four consulting firms to be caught publishing AI-fabricated content, after detection firm GPTZero found hallucinated sources across four PwC Middle East reports in an investigation published this week and verified by the Financial Times. One report claimed a PwC product was used by four national governments, with little public evidence the product exists. Every firm in the world’s most powerful advisory quartet has now shipped the same failure — while selling AI transformation to clients. Here is what was found, how each firm responded, and why India has more exposure to this pattern than almost anywhere else.

What GPTZero found in the four reports

GPTZero’s investigators — Paul Esau, Om Ogale and Alex Cui — ran four PwC Middle East “thought leadership” reports, published between 2024 and 2026, through the company’s AI detector and its Hallucination Check tool, which verifies whether cited authors, titles and URLs actually exist. Their findings read like a field guide to what can go wrong when generative AI drafts prose that no one checks.

The flagship example is a 2025 report titled “Transforming Governance: Citizen Pulse.” It presents a PwC citizen-feedback framework called Citizen Pulse and claims governments in Denmark, Saudi Arabia, the United States and Australia use it. GPTZero checked each country’s cited evidence — Denmark’s Borger.dk portal, Saudi Arabia’s Vision 2030 site, a US Qualtrics press release, Australia’s myGov homepage — and found none of them mention Citizen Pulse at all, concluding there is “little public evidence that the ‘Citizen Pulse’ framework exists outside of this report”. PwC’s own AI-detection score for that document: an 84% chance the report is entirely AI-generated, rising to 100% when the references section is excluded, per GPTZero’s detector.

The other three reports — on Gulf mobility, EV cybersecurity and agentic AI — showed the same pattern in milder form. Across the set, GPTZero documented what it calls “vibe citations”: references that likely came out of a generative model, including a Financial Times-verified example of an academic paper on Riyadh air quality that appears to have been hallucinated entirely, a cited MIT Technology Review article that matches no article the outlet ever published, and a World Economic Forum report GPTZero says does not exist. The FT’s review also surfaced quieter absurdities: a teenage blogger with 280 Medium followers cited as the source on JPMorgan’s AI initiative, and one statistic — that human error causes 90% of traffic accidents — cited three times with three different sources.

Two details turn suspicion into something closer to a receipt. One footnote URL carried the tracking parameter “utm_source=chatgpt.com” — the tag ChatGPT appends when a user copies a link out of its answers. And as Esau put it to the FT: “No human is going to cite the same fact three times in two pages using three different sources.”

PwC’s response: corrections, not a retraction

PwC is not disputing the findings — but it is not withdrawing the reports either. “PwC Middle East takes the accuracy of our published research seriously and is updating a limited number of supporting citations in the reports mentioned,” a spokesperson said, adding that the firm has “quality control processes for research and content development we expect all our people to adhere to” consistent with its Responsible AI approach. As of publication, no retraction has been announced — a notably softer outcome than KPMG’s and EY’s, as the table below shows.

The response lands awkwardly against PwC’s own AI positioning. The firm announced a $1 billion, three-year commitment to AI in April 2023, built around Microsoft’s Azure OpenAI Service, and in May 2024 became OpenAI’s first reseller of ChatGPT Enterprise and the product’s largest user, covering roughly 100,000 US and UK employees, per CNBC. One of the four flagged reports is, per the FT’s description, a playbook for corporate use of agentic AI. The firms selling AI governance are failing their own first test of it: verify what the model wrote before you publish it.

The Big Four scorecard

The PwC story matters most as the completion of a pattern. Since mid-2025, every one of the Big Four has been documented publishing AI-fabricated material — five separate incidents across four firms, with consequences ranging from a government refund to quiet citation edits.

Two things stand out in that table. First, the accountability gradient: a paying government client got money back; readers of free thought leadership get silent edits. Second, the detective is usually the same. The last three exposures — EY in May, KPMG in June, PwC in July — all came from GPTZero, the AI-detection startup founded by Edward Tian and Alex Cui that raised a $10 million Series A in 2024. A startup a fraction of the size of one Big Four office is systematically auditing the auditors, one month at a time — the same cat-and-mouse dynamic we covered in our guide to spotting AI-generated images and videos, now playing out in footnotes instead of pixels.

What it means

The consulting business model makes this failure mode predictable. Thought leadership is marketing: it exists to demonstrate expertise, it is free, and no client signs off on it — so it gets less scrutiny than billable work, while carrying the firm’s full brand authority. Generative models produce exactly the confident, structured prose that format rewards, and they fabricate exactly the things — citations, case studies, statistics — that make it look researched. Detection tools and disclosure norms for this are still maturing across the industry; our AI Tools & Apps coverage tracks both sides of that arms race.

The deeper problem is what these documents are for. Corporate boards and government departments use Big Four reports to justify budgets and policy. When KPMG’s agentic-AI report built case studies — Swiss Federal Railways, Transport for London, an NHS lung-cancer tool — on citations that didn’t check out, the fiction did not stay on the page — it became a data point in someone’s boardroom deck. Fabricated evidence about what AI can do, published by the firms advising on AI adoption, distorts the market’s understanding of the technology itself.

The India angle

No market has more skin in this game than India. The Big Four’s Indian arms and delivery centres are among their largest operations anywhere: Deloitte counts 140,000 people in India and plans 50,000 more hires, EY’s Global Delivery Services runs about 75,000 of its 90,000 people from India — where it opened a Bengaluru AI centre in June as part of a $1.4 billion AI programme — and PwC India employs about 30,000 people, targeting 50,000 by 2030 on FY25 revenue of ₹9,200 crore. A meaningful share of the world’s Big Four research and AI delivery is produced in India — which means India’s professionals are both the engine of this content machine and the ones best placed to fix its verification gap.

India also supplies the clearest warning of where unverified AI text leads. On July 2, the Supreme Court set aside an NCLT judgment that relied on hallucinated, AI-generated case citations — precedents no lawyer had cited, three of which did not exist — calling such fakes “invisible, insidious and catastrophic” and directing the Bar Council to respond. MediaNama has catalogued eleven AI-hallucination episodes in Indian courts since 2023, from tribunals to High Courts. The parallel to consulting is exact: professional documents, institutional authority, invented sources — the same failure we examined when Indian courts weighed how AI systems handle copyrighted text.

What India lacks is a rule. MeitY’s labelling push targets synthetic media on social platforms, SEBI’s proposed AI governance framework covers market intermediaries’ models, and ICAI’s AI committee is focused on adoption and upskilling — as far as we could find, no Indian regulator has issued standards for AI-fabricated content in consulting or audit deliverables, even as Indian outlets like Business Today reported the KPMG withdrawal to an audience of exactly the executives who buy these reports. The Supreme Court has now forced the legal profession to confront hallucinations; the advisory profession’s reckoning, on current evidence, is being outsourced to one small American startup.

What to watch

Three threads from here. Whether PwC’s “citation updates” quietly become corrections pages with change logs — the transparency standard Deloitte Australia was pushed into — or the four reports simply mutate in place. Whether GPTZero’s monthly cadence continues; its investigators say the Big Four series began with government reports and academic papers, and strategy firms like McKinsey and BCG publish the same genre. And whether any client, anywhere, starts writing AI-verification warranties into consulting contracts — the moment fabricated citations stop being an embarrassment and start being a breach. When the firms paid to check everyone else’s homework cannot show their own working, the fix is unlikely to be voluntary.

This article reports on published investigations and company statements; it is not professional advice about any firm’s services.