What is temperature in AI models?

In plain English

At each step, a language model works out probabilities for many possible next tokens. Temperature decides how adventurously it picks among them. At a low temperature it almost always takes the most likely option, producing steady, repetitive, predictable text. Turn it up and less likely options get a real chance, producing more varied and sometimes more inventive writing.

A rough analogy: low temperature is a cautious writer who always reaches for the obvious word; high temperature is one who will try an unusual one and occasionally misfire.

Why it matters

Temperature explains why the same prompt can give different answers on different attempts, which surprises people used to software behaving identically every time. It also matters for anyone building with these models: a tool that extracts data from invoices wants near-zero variation, while a tool brainstorming campaign names wants plenty.

An example

Asked for a tagline for a bakery at a low setting, a model offers “Fresh bread, baked daily” every time. Nudged higher, it produces a dozen different options, some charming, some odd. For the bakery’s opening-hours auto-reply, you would want the boring, consistent version.

What to watch out for

Most consumer chat apps do not expose this control, so adjusting it may not be an option; it is usually found in developer interfaces and settings panels. More importantly, low temperature does not mean accurate. A model can repeat the same wrong fact with perfect consistency, and high temperature does not make it creative in any human sense — it simply widens the range of plausible continuations, including implausible ones.