What is temperature in AI?
A setting that decides how adventurous the model is when it picks each next word. Low keeps it steady; high lets it wander.
Temperature is a setting that shapes how a model chooses each next word. At every step, a model holds odds over the words that could come next; temperature decides how strictly it sticks to the likeliest ones. A low setting makes output steady and repeatable, which suits factual work. A high setting spreads the choices wider, which reads as variety. It changes nothing about what the model knows.
A dial over the odds.
Every step of inference ends the same way: the model has produced a probability for each word in its vocabulary, and something has to pick one. Temperature reshapes that list before the pick. Near zero, the choice is almost always the top candidate, so the same question gets nearly the same answer twice. Raised, the odds flatten, and less likely candidates start winning slots. Nothing is added to or removed from the model's knowledge in either case; only the picking changes.
Low for facts, higher for drafts.
The setting follows the job. Summarizing a paper, pulling claims out of a report, answering questions over retrieved passages: these want low temperature, because you care that the answer matches the sources and not that it surprised you. Brainstorming names, exploring openings for a piece, trying alternate phrasings: these are where a higher temperature earns its keep, since the point is to see wordings the top pick would never give you. Research tools that ground answers in your documents tend to sit at the low end for exactly that reason; the fix for wrong answers is structure and sources, not the dial.
It is not a creativity switch.
The common shorthand says temperature makes a model creative, and the word misleads. Raising it widens the sampling among wordings the model already finds plausible; it does not add knowledge, skill or judgment. Pushed high enough it does the opposite of helping, producing drift and incoherence as unlikely candidates chain together. What reads as creativity comes from what the model learned in training. Temperature only decides how much of that range you let through, much as the context window decides how much of your material it sees at once. The randomness itself has its own page: why AI gives different answers to the same question covers the sampling step this dial lives in.