Why does AI give different answers to the same question?
Because each answer is written by chance, in a controlled way. The randomness is a design choice, not a glitch.
A language model does not look up an answer, it writes one word at a time, and at every step it chooses among candidate words with a weighted roll of the dice. Different rolls, different answers, even from the same question and the same model. The variation lives in that sampling step; the model itself never changes. Lowering the temperature narrows the choices and makes output nearly repeatable when reproducibility matters.
The model computes odds, then rolls.
Running a model on your question produces, at every position, a probability for each possible next word, not a single fixed answer. Something then has to turn those odds into a choice, and that step is a deliberate roll of the dice: likely words are picked most of the time, less likely ones still get picked occasionally. This is called sampling, and the dial that controls it is temperature, which sets how strictly the model sticks to its likeliest options. Two runs with different random rolls produce two different but equally fluent answers, which is exactly what you see when you ask the same thing twice.
Why it is built that way.
Always taking the single likeliest word produces text that is correct-sounding and dead: the same openings, the same phrasings, list answers that circle the same point. A little randomness is what makes prose read naturally and lets a model give you a second formulation when the first did not land. The engineering question is where to set the dial, not whether to have one. For factual research work the dial sits low, because variety is worth little and drift is worth something; for drafting and brainstorming it sits higher, because the second formulation is the point.
What this means for trust.
Variation is not the same as unreliability, but the two are related in one way that matters: a rerun is not a fact-check. Asking again until an answer sounds right selects for phrasing, not truth, and the honest check is the same as it ever was, comparing the claim against a source. That is why grounded research tools treat sampling as secondary: answers get checked against the passages they came from, so it stops mattering which roll of the dice produced the wording. If you need the same output twice, low temperature gets you close; if you need an answer you can rely on, structure beats rerolling.