01 / CLASSIFYClassify the output into four buckets first
Facts can return to material or an authority. Inferences explain facts. Unknowns are questions the material cannot yet answer. Actions change the external world. They cannot receive the same trust. For facts ask where the source is, what the original says and whether date and version match. For inference ask which facts support it and whether another explanation is reasonable. Unknowns should stay explicit. For action check object, scope, permission, preview, receipt and reversibility.
02 / SOURCECheck source version and time
A real but obsolete rule can still cause a wrong answer; a search-result title does not prove its article supports the conclusion. For high-risk work, inspect the primary source, publication or effective date, and whether the answer accurately follows the original. Ask where the conclusion comes from, whether the original can be opened, when it was updated, and whether the system marks missing support as conjecture.
03 / ACTIONVerify that the external world changed correctly
Generating an email differs from sending it. Clicking print differs from holding a complete page. Saving a rule differs from the target website opening. For external action, accept the object, content, state and receipt. “The system says it is finished” is a clue, not completion itself.
04 / CHECKLISTUse a one-minute acceptance checklist
Before submitting or adopting AI output, ask: did it use my material or invented common knowledge; can key facts, figures, dates and citations return to sources; did it state uncertainty too confidently; if wrong, what is the worst consequence and can it be reversed; and have I seen the real result rather than a completion claim? Not every task needs equal strictness, but higher risk leaves less room to skip these questions.
READER COMMENTS
Leave the thought this article gave you.
No comments yet. You can leave the first one.