Part VI: Living With Confident Machines · Page 13 of 14
Use, Verify, or Walk Away
Before relying on an answer, ask two questions: What happens if it is wrong?
How easily can I check it? Together they sort a question into one of four
kinds of response. Try sorting these:
Click a scenario, then click the bin you think it belongs in.
Pick a scenario above, then click the bin you think it belongs in.
A few habits do most of the work of moving a question toward the safe side
of that line:
Ground it
•Paste the source or relevant facts instead of relying on recall, especially for specifics.[36]
•Separate recall from reasoning: give it the numbers and let it compute, rather than asking it to remember them.[7]
Challenge it
•Invite uncertainty: tell it directly to say so when it's not confident.[45]
•Ask again in a fresh chat using different wording: a fact that survives is more trustworthy.[35]
•Avoid leading questions: ask neutrally rather than "isn't it true that X."[26]
Verify it
•Ask for sources and open the one that actually matters, since the citation itself can still be wrong.[39]
•Check names, dates, numbers, quotations, statistics, and cases: that is exactly where fabrication concentrates.[6]
High stakes alone are not the problem. Tested code may be consequential but
easy to check; obscure trivia may be harmless but impossible to verify. The
danger zone is where those two combine: a costly error with no cheap
independent check.
That combination, not the topic or how confident the answer sounds, is
what the two questions above are actually sorting. Match your effort to
where a question lands, from taking it at face value to checking it
against a separate source, and use the habits above to push a
borderline case toward the safer side before you decide.