CodeshareWhen not to trust an AI answer
A practical list, from people who build one. The specific shapes of question where these systems are least reliable, and what to do instead.
We build an AI chat product, which is exactly why this is worth writing. The failure modes are not mysterious, and knowing them makes the tool considerably more useful.
Anything with a precise, checkable value
Specific numbers, dates, version numbers, quotations, citations, legal or medical specifics. These are where confident fabrication is most likely and most costly, because the answer looks exactly like a correct one.
If a number matters, it needs a source you followed — not a source the model named.
Anything recent
A model's knowledge has an edge, and it does not announce where. It will answer a question about last week with the same confidence as one about a decade ago. If recency matters, the answer needs to come from a live search with the links attached.
Anything you asked leadingly
"Why is X better than Y?" almost always produces reasons X is better, whether or not it is. The question smuggled in the conclusion. Asking "compare X and Y" gets you a different and more honest answer.
Anything where being wrong is expensive
Not because the tool is bad at these, but because the cost is asymmetric. Use it to draft, to explore and to check your reasoning; verify before it becomes a decision somebody acts on.
What to do instead
- Ask for the working, not just the answer, and read it.
- Ask for sources and open them. An unopened citation is decoration.
- For anything computational, prefer an answer that was computed over one that was recalled.
- Ask the same question a different way. Answers that survive rephrasing are more likely to be real.
The question is never whether an AI can be wrong. It is whether you would be able to tell.