Generative AI often produces text by assigning probabilities to possible next tokens and selecting from those possibilities. Sampling settings, hidden conversation context, tool results, system instructions, model updates and other runtime details can therefore make two apparently identical prompts produce different answers.
A traditional calculator is expected to return the same result for the same arithmetic expression. A generative language model works differently: it predicts a sequence of tokens, with many possible continuations receiving different probabilities.
More than one continuation can fit
Suppose a sentence could naturally continue in several ways. The model does not necessarily store one fixed sentence that must be returned. It estimates which tokens are plausible given the text already present. Systems can then choose the highest-probability continuation or sample among several likely choices. Settings commonly described with terms such as temperature or top-p can change how narrow or broad that selection is.
Small differences early in generation can cascade. Once one token is chosen instead of another, the next prediction is conditioned on a slightly different sequence, so two answers can gradually move in different directions while both remain relevant to the original question.
The prompt may not be the whole context
What a user sees in the message box is not always the only information the model receives. A chat system may include earlier messages, system instructions, retrieved documents, tool outputs or application-specific context. If any of those differ, the effective input differs too. Even whitespace, formatting or an updated conversation history can sometimes alter later predictions.
Services also change over time. Providers may update model versions, safety systems, retrieval methods or default generation settings. That means an answer repeated weeks later may not be produced under exactly the same conditions even if the visible wording is identical.
Variation is not the same as unreliability
Output variation can be useful for brainstorming, rewriting and open-ended tasks because there may be many good answers. For factual work, however, variation is a reminder that fluent wording is not proof of correctness. Important claims should still be checked against reliable sources.
Some systems can be configured for more deterministic behavior, but perfect repeatability is not guaranteed across every product and setup. The simplest explanation is that generative AI usually operates over a distribution of plausible continuations, while the surrounding software decides how those possibilities are selected and what context is supplied.