The single most dangerous thing about a good AI model is how convincing it sounds when it is wrong. It does not hedge, it does not sweat, it just states the confident falsehood in the same tone as the truth. Here is how to catch that before it costs you something, without turning every answer into a research project.
Know where it lies
Models are most likely to invent things in predictable places: specific facts and figures, names and dates, quotes, citations, legal or medical specifics, and anything very recent. They are most reliable on general explanation, reasoning about text you provided, and rephrasing. Calibrate your suspicion to the task. A summary of a document you pasted is usually safe. A confident statistic with a source you did not give it is a red flag until proven otherwise.
The three checks that catch most errors
- Ask for the source, then check the source. If it cites something, verify the citation actually exists and says what the model claims. Fabricated references are the classic tell.
- Cross-examine with a second model. Ask a different model the same question. When two independent models disagree, at least one is wrong, and you have found your soft spot.
- Make it show its work. Ask it to explain the reasoning step by step. Confident nonsense often falls apart the moment you ask how it got there.
Treat an AI like a brilliant intern with no shame. Useful, fast, and absolutely capable of handing you a made-up number with a straight face.
Match the effort to the stakes
You do not need to fact-check a brainstorm or a first draft. You absolutely need to check anything going to a client, into production, into a legal document, or in front of your boss. The rule is simple: the closer the output gets to a real-world consequence, the harder you check.
The mindset
Use the model as a fast first draft and a thinking partner, never as a final authority. It is right often enough to be genuinely useful and wrong often enough to hurt you if you stop paying attention. Trust it the way you trust a smart, tired colleague: gratefully, and with your eyes open.