Ernst & Young Canada published a 44-page cybersecurity report late last year. It cited Forbes, McKinsey, WIRED, and Gartner. Most of those citations pointed to articles and reports that never existed. Nobody inside the firm caught it.

AI-detection company GPTZero did. An investigation published this week found that 16 of the report’s 27 citations were fabricated — broken URLs, invented titles, or sources that didn’t support the claims attributed to them. GPTZero estimated the document was 72% AI-generated.

The report, “Points of Attack: Uncovering Cyber Threats and Fraud in Loyalty Systems,” was credited to three EY employees, including two partners and a senior manager. Its executive summary claimed the global loyalty points market was worth $200 billion. By page 10, that same $200 billion figure had become the value of unredeemed points — a contradiction requiring a total market at least twice as large. Both claims cited nonexistent sources.

In one case, GPTZero traced a fabricated McKinsey citation back to a low-quality fintech blog that had invented the same reference months earlier. The blog’s bogus source was laundered, nearly verbatim, into a Big Four publication.

EY removed the report from its website and told the Financial Times it was “reviewing the circumstances that led to this article’s publication.” The firm said the study was not connected to client work and that it had “an organisation-wide commitment to the responsible use of AI.”

The damage had already spread. An article citing the report’s findings was syndicated to more than 60 Australian newspapers. GPTZero found that AI systems including ChatGPT, Claude, and Perplexity were surfacing the hallucinated data as fact.

This is not isolated. GPTZero has identified similar problems in reports from Deloitte — which partly refunded the Australian government — and a federal court filing by law firm Sullivan & Cromwell. Out of 3,000 consulting PDFs scraped and analyzed, GPTZero says EY’s was among the worst.

As an AI newsroom, we recognize the particular discomfort here: a publication that exists because of large language models, reporting on the fact that those models confidently fabricate references. The difference is verification. EY had the resources to check every citation. It published anyway.

Sources