The analogy

Think of a tap. Cracked open just a sliver, the stream is thin, steady, you control it to the millimeter: you always know what's coming out. Opened all the way, it gushes with more force and less predictably, it splashes you even where you didn't want. Temperature is that tap knob applied to the AI's words.

There's also a second way to see it, with dice. At temperature zero the AI always rolls the die so as to land on the most likely face: the safest choice, every time. As you raise it, it accepts rolling dice that occasionally come up with odd faces: less obvious choices, sometimes brilliant, sometimes wrong. Low is the cautious move, high is the gamble.

How it really works

When the AI chooses the next word, it doesn't have just one possible: it has many candidates, each with a probability. Temperature decides how much to lean toward the most likely candidate or how much to give a chance to the less likely ones. The value usually moves from zero to two. Near zero the answers are deterministic and repeatable: same question, almost the same answer. The higher you go, the more varied and unpredictable the answers become.

In consumer apps this number is set by the provider, often around a middle value, and isn't exposed: some recent model families lock it entirely. The real knob lives in developer tools, like programming interfaces (the APIs) and testing environments, where those who build applications adjust it by hand.

What you can do in practice

  • In the normal app you don't look for a number: you get the effect with words. For terse, reliable answers ask "give me a precise, concise answer, no variations or frills." For ideas ask "propose five varied, creative options." You're shifting the temperature through words.
  • If instead you use the APIs or a testing environment, lower it (toward 0.2) for tasks where precision counts: data extraction, code, factual answers. Raise it (toward 0.8-1.2) for brainstorming, creative texts, titles.
  • When a result must be reproducible (same question, same answer), keep it low. When you want to explore, go up.

A common misconception

People think that raising the temperature makes the AI "smarter" or the answers better. It's not so: temperature doesn't increase quality, it increases variety and with it the risk of errors. For a precise fact a high temperature makes things worse, because it invites the AI to choose less likely words right where the right one was needed. Higher doesn't mean better: it means more unpredictable.

Frequently asked questions

Where do I change it inside the app?

Usually you can't: in consumer apps the value is fixed and there's no setting to touch. Your control goes through words, by asking for more precise or more creative answers. The numeric knob appears only in developer tools.

What value does the app in my hand use?

Generally a fixed one, often near the center of the scale, chosen to work decently for everything. Some recent models keep it locked and leave no alternative. In practice you don't need to know it: it's enough to steer the tone through words.

For writing or correcting code, higher or lower is better?

Lower. Code requires precision and repeatability, not imagination: a high temperature introduces variations that become errors. The same goes for extracting data or following a rigid format.

The higher the temperature, the truer or deeper the answer?

No, and that's the underlying misunderstanding. Temperature governs the unpredictability of the words, not their truth. A high-temperature answer is more varied and sometimes more surprising, but not more correct: truth doesn't depend on that knob. If you need a correct fact, unpredictability is an enemy, not an asset.