The difficult part of using an AI assistant is not that it makes mistakes. Everything makes mistakes. It is that a wrong answer arrives looking exactly like a right one: same fluent sentences, same steady tone, same air of authority. There is no tell in the writing itself.
So you cannot rely on how an answer feels. You need to know where errors tend to appear, and check those places specifically.
Where mistakes cluster
Anything with a number in it
Statistics, dates, prices, version numbers, distances, dosages. Numbers are exactly the kind of detail a language model reconstructs rather than recalls, and a plausible number is easy to produce. Treat every specific figure as unverified until you have seen it somewhere else.
Citations and sources
A reference with a real-sounding author, a plausible journal and a convincing year may still not exist. This is the single most damaging failure mode, because a citation looks like evidence. If a source matters, open it. If it will not open, it may never have existed.
Recent events
Models are trained up to a cutoff date. Ask about something newer and you may get a confident answer built from older information, unless the assistant explicitly searched. Check whether it actually looked something up or simply answered.
Anything specific to you
Your company's policy, your local law, your particular device model. The model knows the general shape of such things and will fill in the specifics plausibly. General shape plus invented specifics is a dangerous combination when you are about to act on it.
A thirty-second check
You do not need to fact-check everything. Scale the check to the consequence of being wrong.
- Ask yourself what happens if this is wrong. If the answer is 'nothing much', move on.
- Identify the load-bearing claims: the two or three facts the rest depends on.
- Check those against a primary source, not another AI.
- Ask the assistant directly: 'which parts of this are you least confident about?' The answer is often revealing.
- For anything legal, medical, or financial, treat the output as a starting point for a question to a professional, never as the answer.
Prompting that reduces invention
- Supply the source material rather than relying on recall. A model summarising your pasted document invents far less than one answering from memory.
- Give explicit permission to say 'I don't know'. Models default to producing an answer; telling them uncertainty is acceptable makes it more likely.
- Ask for the reasoning, not just the conclusion. Faulty logic is much easier to spot than a faulty fact.
- Ask the same question twice in different words. Answers that change substantially indicate the model is guessing.
The habit worth building
Treat an AI answer the way you would treat a confident summary from a well-read colleague who has not checked their notes: a genuinely useful starting point, worth verifying before you repeat it or act on it. That posture gets you most of the speed benefit with very little of the risk.
AI Guide has a full learning path on research and verification workflows, covering how to check sources and structure prompts that reduce invented detail.