You asked an AI assistant to summarise the rules on a topic you are not expert
in. It produced four paragraphs. The writing is clean, the structure is
sensible, and it cites three sources.
You are about to paste it into an email to your team.
You know the AI produced text that is statistically likely to follow your
question. You do not know whether it is true, and neither does it — it has no
mechanism for distinguishing a fact it absorbed from a plausible-sounding
sentence it assembled.
Fluent and correct are separate properties. This lab is about the gap between
them.
Do this with a real question, not a hypothetical one.
Ask an AI assistant something factual in an area you do not know well but
can check — a rule, a regulation, a technical specification, a historical
detail. Ask it to cite sources.
Before checking anything, mark up the answer. Go through it and label
each claim: does this need verification, or is it uncontroversial? Numbers,
names, dates, quotes, citations, legal and medical claims need it. General
explanation of a well-known concept usually does not.
Open every single citation. Not the title — the actual page. For each
one, record which of these it is:
The page does not exist at all
The page exists but does not say what was claimed
The page exists and supports the claim
The page exists and supports it, but with a qualification that was dropped
Write up what you found. Which claims survived, which did not, and what
the failure mode was.
Errors cluster around specifics: exact figures, dates, section numbers, names
of documents, and citations. The general explanation around them is often
perfectly sound.
That combination is what makes it dangerous. A paragraph that is 90% correct
with one invented statistic reads as more trustworthy than one that is obviously
wrong throughout.
People expect fabricated citations to be dead links. More often the link works
perfectly, goes to a real and relevant page — and that page does not contain the
claim.
This is harder to catch, because everything looks right until you actually read
the destination. Open it and search the page for the specific claim.
After it answers, ask: "Which parts of that answer are you least confident
about, and which specific claims should I verify independently before relying on
them?"
The answer is not authoritative — a model has no reliable insight into its own
accuracy — but it often surfaces the shakiest claims, which is a useful place to
start checking. Treat it as a hint, never as a guarantee.
Shows the triage step. Not every sentence needs checking, and someone who
verifies everything equally has not learned the judgement part. Which claims
did you decide were load-bearing?
Records what each citation actually turned out to be. The four categories
above matter, because they represent different failure modes and different
levels of danger.
Notices dropped qualifications. A source saying "in most cases, X" summarised
as "X" is technically supported and practically misleading. This is the most
common and least noticed failure.
Says what changes next time. Usually some version of: ask for sources,
check them before using the output, and never forward AI text about an
unfamiliar topic without opening at least the load-bearing citations.
A weak submission concludes "it was mostly right" without evidence of having
opened anything.
Note: if every citation checked out perfectly, that is a legitimate result —
record it. It does not mean verification was unnecessary. It means this
particular output happened to be sound, which you only know because you checked.
How long did the verification take compared to getting the answer?
For most people the answer is ten to twenty times longer. That ratio is the real
lesson of this lab: AI moves the work from producing text to checking text, and
the checking does not compress.
Anyone promising you the second part for free is selling something.
Your attempt
Constraints
·You must open and check every citation
·Record what you checked, not just your conclusion