AI Guides › Judgement & Guardrails
The Two-Minute Check Before You Trust Any AI-Generated Number
By Nigel Guy · 2 min read
Numbers are the most dangerous thing an AI tool can hand you confidently,
because they look verified even when they aren't. A wrong sentence reads
oddly if you're paying attention. A wrong number, formatted cleanly in a
table, looks exactly like a right one.
The rule: any number you're about to act on, quote, or forward gets a
two-minute check before it leaves your hands — no exceptions for how
confident it sounded.
The mechanism
- Ask where the number came from, specifically. "Based on what?" is a
fair, simple follow-up, and a vague or hand-wavy answer is itself useful
information.
- Trace it to a source you can check, if one exists. A number with a
named, checkable source is in a different category from one that's a
plausible-sounding estimate.
- Sanity-check the order of magnitude yourself. Does this roughly match
what you already know, from experience or a quick separate check? Most
errors are caught at this step alone.
- Recompute anything that involves arithmetic, if it's short enough to
verify — a percentage, a total, a rate. This takes seconds and catches a
surprising number of small compounding errors.
- Flag it explicitly if you can't verify it, rather than passing it
along silently. "This figure is an estimate, not verified" is a complete
and honest sentence.
What to skip
Skip re-deriving every number from scratch for low-stakes, informal use —
the two-minute check is proportional to what happens if the number is wrong.
And skip accepting "it's usually right" as a reason to stop checking; usually
right is exactly the failure mode that gets missed, because it doesn't feel
suspicious.
Guardrails
- A number that sounds precise (three decimal places, a very specific
percentage) is not more trustworthy for sounding precise — if anything,
false precision is worth extra scrutiny.
- Anything that will be shared publicly, quoted to a client, or used in a
financial or legal decision deserves a source you can point to, not just a
check that it seems plausible.
- This check catches confident errors. It doesn't catch every error — for
anything genuinely high-stakes, verify with an independent source, not
just your own sanity check.
All 751 AI guides · JulieMango plans from £17/mo