AI Guides › Judgement & Guardrails
Reading An AI Citation Like You'd Read A Stranger's Homework
By Nigel Guy · 2 min read
A citation that looks right — author name, title, year, even a page number —
reads as verified. It has the shape of something someone checked. Nothing
about that shape actually tells you the source exists, says what's claimed,
or has been quoted accurately. The formatting and the accuracy are entirely
separate things, and only one of them is easy to fake convincingly.
The rule: treat every AI-generated citation as an unverified claim of
existence until you've found the source yourself — the citation format
proves nothing on its own.
The mechanism: the homework check
Mark a stranger's homework, and you don't take the working on trust just
because it's written neatly. Apply the same standard here.
- Does the source exist at all? Search for it directly — title,
author, publication. An invented source usually turns up nothing, or
turns up something with a similar but not identical name.
- Does it say what's being attributed to it? A real source cited for a
claim it doesn't actually support is arguably worse than a fabricated
one, because the existence check alone won't catch it.
- Is the quote, if there is one, exact? A paraphrase presented as a
direct quote is a real category of error, distinct from outright
fabrication, and needs its own check — read the actual passage, not just
confirm the source is real.
- Is this the current or correct version? A source that existed once
but has since been revised, retracted, or superseded can be technically
real and still misleading to cite as current.
- Would this citation survive someone else checking it? If you're
about to hand it to a reader, a client, or an editor, assume they might
check, and make sure it holds up when they do.
What to skip
Skip this level of verification for citations used only in early drafting
or brainstorming, where you're exploring what a source might say before
committing to using it. And skip assuming a citation is safe simply because
it names a well-known publication or author — a real name attached to a
fabricated title or claim is a common shape for this error to take, not a
rare edge case.
Guardrails
- Anything that will be published, submitted, or cited to someone else
needs every citation checked against the actual source, not a sample of
them.
- A source you can't find after a genuine search attempt should be treated
as unverified, not as "probably real but obscure." The honest label is
"I couldn't confirm this," stated plainly.
- This applies to your own citations too, if you're drafting with AI
assistance and adding sources from memory — memory fabricates with the
same fluent confidence a model does, and deserves the same check.
All 751 AI guides · JulieMango plans from £17/mo