Verifying AI output

"Always verify everything" is advice nobody follows, because it costs more than the tool saves. What works is a routine where the depth of verification follows the stakes.

The three tiers

Tier 1: skim (internal, low stakes)

Brainstorms, internal drafts, reformatting your own material. Read it once for obvious nonsense and tone. The facts are yours; the model only shaped them.

Tier 2: fact-check the load-bearing points (work that travels)

Anything leaving your hands: reports, client emails, presentations. Identify the load-bearing claims (the numbers, names, dates, and quotes the document depends on) and verify each against an independent source: the original document, the official website, the primary data. If the model provided sources, open them; models can cite real pages that do not say what is claimed.

Tier 3: expert review (consequential or specialized)

Legal wording, financial figures, medical content, security configurations, regulatory statements. AI may draft; a qualified human owns the result. The review is not optional politeness, it is the control that makes AI use defensible.

Finally, watch for automation complacency: the documented tendency to review the first outputs carefully, find them good, and gradually stop reviewing at all. The hundredth AI draft deserves the same tier as the first, because the model did not get more reliable; you just got more comfortable.