LearnGrok"Always verify everything" is advice nobody follows, because it costs more than the assistant saves. A proportionate habit is one you will actually keep.
Sort the output by how it fails
Things that fail loudly. Code that doesn't compile, JSON that doesn't parse, a script that errors. You do not need to review these carefully, because running them is the check. Spend your attention elsewhere.
Things that fail silently. Numbers, dates, names, citations, quotes, attributions, legal and medical specifics, and anything stated with more precision than the question deserved. These are where a wrong answer looks exactly like a right one. Check every single one you plan to repeat to another person.
Things that cannot really be wrong. Rewrites, tone changes, brainstorming, "give me ten options". Skim and move on.
Three checks that catch most of it
- Ask for the source, then open it. Not "did you make this up?": models are unreliable narrators of their own confidence. Ask where it came from and click through. A link that doesn't exist, or exists and doesn't say that, is the answer.
- Re-ask in a fresh conversation. Same question, no context from the first attempt. Two independent runs that disagree on a fact means neither is trustworthy.
- Check anything time-sensitive against the source of truth. Prices, limits, version numbers, availability, "is X supported yet". A model's training data has a cutoff, and even with live search the answer may be assembled from stale pages. For anything about Grok itself, that source of truth is x.ai and docs.x.ai.
The failure mode to watch for in yourself
The risk is not that you believe a bad answer once. It is that after fifty good answers you stop looking. Attach the check to the type of claim rather than to your confidence in the moment, and it keeps working when you are tired.
What changes
Which model backs your conversation, and whether it searched the web for a given answer; both affect how much checking a claim deserves, and both change without notice. The sorting above is deliberately independent of that.
