Forty-one links. Two questions.

Guy-PT hands over a training program with forty-one sources stapled to the back of it. The first link Taylor picks is dead. So is the second. Claudio opens a third and finds the same thing, and then says what everybody says at that point — thirty-eight left, that doesn’t scale — and the episode just… ends. That’s not me being coy. It’s that the answer doesn’t fit in a speech bubble, and I didn’t want to fake one.

The method is two questions, not one

Most people treat “check the citations” as one job. It’s two, and they cost wildly different amounts. Separating them is the whole trick.

Run question one across the whole list before you run question two on any of it. Question one is uniform, mindless, and fast, and it thins the pile. Question two is where the time actually goes, and you want to spend it on a list that has already survived something.

Across the eight reference rows we captured from GPTZero’s table, the same verdict comes back four times — “URL returns a 404 error.”⁴ A fifth says only that the URL is broken. That is GPTZero’s annotation, not EY’s text — somebody had to go and click. The cheapest question in the world.

What the first link actually tells you

It is a gate, not a sample. This matters and it’s the part people get backwards.

If the first link is broken, you have learned something about the whole set — stop trusting it, you’re holding a decoration rather than a document. If it holds, you have earned the right to keep checking, not the right to stop. A trainer watching your first rep isn’t doing statistics. They’re deciding whether the session happens.

Stop trusting isn’t stop checking, though, and that’s the part I’d get wrong reading this fast. You still work the list. In the comic they open three, find all three broken, and then sit down with the other thirty-eight anyway — which is the least dramatic thing that happens in it.

What to do with the dead ones

A dead link kills the citation. It doesn’t automatically kill the point the citation was holding up — those are two separate calls. Strike the citation, then go back to the sentence and ask what’s left standing. Usually the general claim survives and the specific number doesn’t. If you can’t re-ground the number, the number doesn’t get to keep travelling with the claim.

I’m not going to tell you how long forty-one links takes, because it depends entirely on the list and anybody who gives you a number is guessing at your afternoon. What I’ll say is that the expensive-sounding objection usually isn’t about time. In the comic Taylor read the whole program on a Sunday and never clicked once, and his reason was that he liked how thick it was. I have done that. Not with a training program.

Don’t take my word for it

The evidence slide in the episode makes three statements that aren’t jokes. Here’s exactly where each one comes from, quoted and linked. Click them. Hand them to your AI and let it check my numbers, too — that’s the entire point of the episode, and it would be a strange page to end with “trust me.” (There is a fourth factual thing in the comic — the number on Taylor’s screen. It gets its own section, for reasons that will become obvious.)

That last row used to say something else, which brings me to the part I’d rather not put on a public page.

Two things I nearly printed

Both got caught before publication. The first one by a person re-reading a number that looked too clean; the second only because somebody went and opened the actual page instead of trusting the file we’d already built from it. I’m including them because a page about checking citations that presented itself as never having needed correcting would be its own kind of joke.

It wasn’t only EY

In 2025, Deloitte Australia delivered a report to the federal Department of Employment and Workplace Relations. A University of Sydney researcher, Chris Rudge, started reading it closely and found it “littered with citation errors”⁸ — including “a fabricated quote from a federal court judgment and references to nonexistent academic research papers”⁹. Deloitte agreed to partially refund the A$440,000 the government had paid.⁹

Here is the detail that matters for question two. The invented references were attributed to real academics — researchers who exist, at universities that exist, “neither of whom wrote what the report claimed they wrote.”¹⁰ Question one would have sailed straight past that. Real name, plausible title, nothing obviously broken. It takes somebody actually opening the thing and reading for the claim.

The EY report has a version of the same move: GPTZero traced a fabricated McKinsey citation¹⁴ back to a blog post, where the same invented source already appeared — “laundering an invented source from a low-quality blog into a Big Four publication.”¹¹ A citation can be perfectly formatted, confidently repeated, and load-bearing for a claim, and still be about nothing.

Neither firm is the villain of the comic. The failure mode is simpler than that: nobody checked the citations against the sources. At EY that surfaced as dead links; at Deloitte as real scholars who never wrote the work they were credited with. Both were caught from outside — a research team at GPTZero, and one academic reading closely. Neither firm’s own process caught it.

Why a hiring manager already cares

Guy-PT spends the episode standing on a credential nobody ever verified, which is a joke until you look at what the people doing the hiring say they’re screening for. In Korn Ferry’s talent-acquisition trends report for 2026, “nearly three-fourths (73%) list critical thinking as their No. 1 priority when evaluating potential hires, while AI-related skills placed fifth.”¹²

One of their own experts puts the connection more bluntly than I would have dared: “You need critical thinking to understand what’s a hallucination versus real data.”¹³

Nothing on this page is legal, compliance, medical or other professional advice; it’s a comic with its homework shown. The sources are all below — go read them, or point your AI at them and see whether it agrees with me.