EY Got Caught Filing a Cybersecurity Report Full of Hallucinated Citations

A Big Four logo on a PDF used to mean something. When EY, Deloitte, PwC, or KPMG ships a six-figure report to a government agency, the contents get treated as authoritative on contact. That assumption just took another serious hit — this time from EY Canada, on a cybersecurity advisory engagement of all things.

Most of the citations pointed to documents that don’t exist

Researchers vetting an EY cybersecurity report submitted to a Canadian government body started doing something deceptively simple: checking the footnotes. What they found was that a large share of the cited academic papers, government documents, and references did not exist.

Fake authors. Fake journals. Page numbers pointing to nothing. The titles looked plausible, the years looked plausible, the formatting was clean — but search for any of it and you get zero hits. This is the textbook signature of generative AI hallucination, the same pattern that’s been embarrassing lawyers in US courtrooms for two years now.

The uncomfortable part: this wasn’t an internal memo or a draft deck. It was a formal advisory document positioned to shape government cybersecurity policy.

This keeps happening, and the pattern is the tell

EY Canada is not an outlier. In Australia, Deloitte had to issue a partial refund on a AU$440,000 welfare system report to the federal government after researchers found fabricated citations and a quote attributed to a federal court judgment that didn’t exist. US federal courts have sanctioned multiple attorneys for filing ChatGPT-generated briefs citing imaginary case law — the Mata v. Avianca case in 2023 was just the opening act.

The pattern is consistent. The more expensive and prestigious the deliverable, the more incentive there is to hide the AI use, and the weaker the internal review becomes. A consultant billing $800 an hour is not going to volunteer that they pasted a prompt into ChatGPT and shipped the output. So the verification layer that justified the price tag in the first place quietly disappears.

What clients are actually paying for

Step back and ask the obvious question. Why do governments and Fortune 500s pay Big Four rates in the first place? It was never really about the report itself — McKinsey-style decks are not exactly a scarce resource in 2026. The product is the verified expertise and accountability behind the logo.

Brand-warrantied facts. Deep institutional read of industry practice. A reputation big enough to sue if things go sideways. Strip that out and a Big Four firm becomes something much less interesting: a high-priced reseller of LLM output you could have generated yourself for $20 a month.

The EY Canada incident isn’t really a story about one bad report. It’s a signal that the review workflow itself is broken. Senior partners are signing off on documents where no one checked a single footnote.

When AI slop puts on a tie

Here’s the genuinely dangerous part. When a random Substack writer publishes AI-generated nonsense, readers apply some skepticism. When a PDF with the EY letterhead lands on a policymaker’s desk, it goes straight into the citation graph. That document gets quoted in the next report, which gets quoted in the one after, and you end up with a policy ecosystem built on sources nobody can actually find.

This is why “AI slop” stopped being a meme about bad Facebook images and started being a structural concern. The slop itself isn’t new. What’s new is that it now travels with institutional credentials attached. The EY case is evidence that the firewall between “stuff the internet made up” and “stuff trusted institutions vouched for” has already been breached.

The takeaway

Expert authority doesn’t come from the job title, the firm name, or the page count. It comes from someone, somewhere, having actually checked the facts and being willing to stake their reputation on them. As long as AI lets that step get skipped, every Big Four logo is going to draw more suspicion, not less.

Next time a hundred-page consulting deliverable lands in your inbox, try one question: “Who verified these citations?” If the answer involves a pause, you already have your answer.