Why two AI answers to the same question differ

Two answers to the same visible question can differ without the system having discovered a deeper truth on the second attempt. One possible reason is how the next token is selected.

Google’s model-configuration guidance describes temperature as controlling randomness in token selection. Top-P and top-K restrict the candidate tokens considered. A seed can support repeatability, but the documentation does not promise perfectly deterministic output. At temperature zero, the highest-probability choice is mostly deterministic, with some variation still possible, and parameter availability differs by model. Model or parameter changes can also change the result.

Those mechanics explain one possible source of variation. They do not tell us why two unknown answers from a particular interface differed, because the visible question may sit inside different context or settings. They also do not imply that a lower-randomness answer is more correct.

Here is a fully fictional example. A user asks, “Which train leaves first?” The supplied timetable says Train A leaves at 9:10 and Train B at 9:25. One answer says, “Train A leaves first.” Another says, “Train B leaves first.” The wording differs in one case, but the substantive claims conflict. Repeating the question until one answer appears more often does not resolve the timetable. The timetable does.

A different pair could say the same thing in different words: “Train A departs first” and “A is earlier, at 9:10.” That is variation in wording, not a disagreement about the fact. The first task is to classify the difference.

My practical habit is to preserve the question, supplied context, relevant settings when known, and source. Then I compare the substantive claim with the source. If the answers differ only in phrasing, choose the clearer one. If they differ in facts, inspect the evidence. If no source exists, mark the uncertainty instead of voting among repetitions.

A second answer can be useful as a prompt to check the claim. It is not automatically a second witness. The configuration guide explains how variation can arise; it does not turn repetition into verification. The reader still has to decide which part of the answer can be checked and where the evidence lives.