How to read an AI answer like an auditor

Auditors read differently from the rest of us. They do not ask “does this sound right?” — they ask “what would I have to see for this to be right, and is it here?”

That habit transfers directly to reading AI output, and almost nobody has been taught it. So here is the discipline, compressed to five checks. They take about two minutes. They are worth it for any answer you intend to act on.

The five checks

1. Find the load-bearing claim. Every analysis rests on one or two assertions that, if wrong, collapse the conclusion. Find them before judging anything else. If you can’t tell which claims are load-bearing, the answer is structured to impress rather than to inform — that is itself a finding.

2. Check whether the load-bearing claims carry sources you can open. Not “studies show.” A named source, with a link or citation you could follow now. The research on model hallucination is consistent: specific names, numbers, and citations are what models most confidently invent. An unsourced specific is an unverified specific.

3. Ask what was assumed. Every recommendation assumes things — about your context, your constraints, what “success” means. A trustworthy answer states them. If it doesn’t, state them yourself: “this assumes my team is office-based; it isn’t” kills more bad recommendations than any fact-check.

4. Look for the uncertainty — and be suspicious if there is none. Real analysis of a real question always has parts that are less sure. If everything is delivered at the same confident register, the uncertainty hasn’t disappeared; it has been hidden from you, and you are carrying it unknowingly.

5. Identify who stands behind it. If you act on this and it is wrong, who — besides you — answers for it? A named reviewer changes the answer’s nature. If the honest answer is “no one,” then you are not receiving advice. You are receiving raw material, and you are the reviewer now.

The point

Notice that none of the five checks requires technical knowledge of AI. They are the same questions a careful board member asks of a consultant’s deck, or an editor asks of a reporter’s draft. Machines have not changed what diligence is; they have changed how much confident, fluent, unverified material arrives per day — which makes the old discipline more valuable, not less.

We built these checks into the Standard because we think answers should arrive already carrying them. Until that is normal, carry the checklist yourself.