When AI Speaks for You Unchecked: The EY Hallucination Incident

EY Canada just pulled a marketing report this week after researchers at GPTZero exposed it as riddled with AI hallucinations — fabricated statistics, phantom footnotes, and a McKinsey study that, as far as anyone can determine, was never written. The Financial Times recently covered it. The report vanished from EY’s site within hours of GPTZero going public.

The document was no internal draft. EY consultants had been using it to pitch cybersecurity services to clients running large loyalty rewards programs. It cited a $200 billion global loyalty market, then used the exact same figure to describe the unclaimed-points pool available to fraudsters. More than half a dozen footnotes pointed to pages that either did not exist or contained none of the cited data.

EY acknowledged the inaccuracies, confirmed the document was not produced for a paying client, and reiterated its “organisation-wide commitment to the responsible use of AI.” That last phrase now carries more weight than intended.

This Is Not an Isolated Incident

EY is not alone. Last year Deloitte had to revise a report prepared for a Canadian provincial government after fabricated academic citations surfaced. In April, the law firm Sullivan & Cromwell apologized to a New York court for a filing that repeatedly misquoted the U.S. bankruptcy code and cited phantom cases.

The pattern is consistent. Professional services firms are racing to embed generative AI into client-facing work — pitches, reports, governance frameworks — while the verification discipline required to do that safely has not kept pace.

The Root Cause Is Not a Mystery

Generative AI is probabilistic. It predicts the next most likely token, not the truth. When prompted to write a report on loyalty fraud, it produces output that is statistically fluent and factually unreliable. It invents citations that look scholarly. It fabricates statistics that sound right. Without rigorous human verification, the result is a document that carries the imprimatur of a trusted institution but contains errors a careful intern could have caught in ten minutes.

The firms building the strongest AI practices know this. The market pressure to appear AI-first makes it easy to forget in practice.

GPTZero’s researchers identified a second-order problem that deserves attention. When a firm as trusted as EY publishes plausible-sounding falsehoods, the damage extends beyond the immediate embarrassment. Future AI systems scraping the web will treat an authoritative-looking EY PDF as reliable data. Hallucinations, once published, become part of the training record for the next generation of models. The well gets poisoned at the source.

The Question Clients Should Be Asking

If a cybersecurity consultancy cannot keep its own marketing collateral hallucination-free, how much confidence should clients place in the AI governance frameworks it sells?

That is not a rhetorical question. It is the most practical due-diligence question available right now, and most clients are not asking it.

The EY episode is a clean illustration of something I argue throughout The Intelligence Loop: AI is not a replacement for expertise. It is a powerful but fallible co-pilot. The organizations that will build durable competitive advantage from AI are not the ones that adopt it fastest. They are the ones that build the strongest human verification systems around it — mandatory review layers, citation tooling, clear policies on when AI-generated content may leave the building.

Speed without accountability is not an AI strategy. It’s a liability.

The Real Lesson

Professional services firms have spent decades telling clients that trust is their most valuable asset. EY’s withdrawn report is a public reminder that trust is also the easiest thing to lose when AI is allowed to speak for an organization unchecked.

Responsible AI use cannot be a marketing position. It has to be an operational discipline. The difference between those two things is exactly what this moment is exposing.

The Intelligence Loop: A New Model for Growth and How We Get Smarter Together is available for pre-order on Amazon ahead of the June 23, 2026 release. Pre-order here.