Why AI invents things instead of saying “I don't know”
A language model does not look up answers in a database of facts. It continues text the way such text usually continues, and it does so very convincingly. If you ask for a specific number or link and the model has no reliable data, it will still produce something plausible: the right format, a confident tone and nice details. That is why you cannot tell truth from invention by how it sounds. A smooth, detailed and calm answer says nothing about whether it is correct. Remember this: doubt should switch on not when an answer sounds strange, but when it contains something checkable and something depends on it.
What AI invents most often
Small, checkable details are the most dangerous. The most frequently invented items are exact numbers and percentages, dates of events, names of people and organisations, law and article numbers, quotes “from a book” or “from an interview”, web addresses and titles of research papers, and program functions or settings that do not exist. A link to a source is not proof either: AI can invent those too, complete with a believable publication name and year. General reasoning, explanations of concepts and rewording your own text are invented much less often, so that is where to point the AI first.
Four checks that take five minutes
First: ask for the source and open it yourself. If the page does not exist or says something else, do not use the answer. Second: find the original value. Check a figure on the organisation's official site, search a quote by its exact words in quotation marks, verify a law on your country's official portal. Third: ask the same question in a new chat in different words. If the answers differ noticeably, that is a warning sign. Fourth: ask the AI to split its answer into “confident” and “not sure, needs checking”. It is no guarantee, but it shows weak spots. The main rule: anything you are going to publish, send or sign is checked against the primary source.
How to write a prompt that reduces made-up answers
The best way to cut invention is not to ask the AI to recall, but to give it material. Paste a text, a table or an excerpt and say that it must answer only from that. Add permission to say “the text does not say”: many models need to be told explicitly that not knowing is allowed. Ask for short answers with a quote from your material, so each claim is easy to check. Do not demand “exact figures from memory” and do not push when the AI hesitates: pressure raises the chance of invention. If you need fresh data, use a mode with web search where one exists, and still open the links yourself.
When you must not trust AI without checking
Some topics carry a high cost of error: health and medicines, legal questions and deadlines, taxes and money, safety, and anything you publish under your name or hand to a client. In these areas AI is fine as a helper for preparing questions and a draft, but not as the source of a decision. For health and law the final answer comes from a specialist, and numbers and links are confirmed by official sources. A useful habit: before using a result, ask yourself what happens if it is untrue. If the answer is “nothing serious”, a quick check is enough. If “there will be consequences”, check the primary source or consult a specialist.