Temperature is a setting that controls how random an AI model’s text output is. Turn it down and the model mostly picks the highest-probability word, so the same question gives you nearly the same answer every time; turn it up and lower-probability words get a chance, so replies grow more varied but also drift and err more often. It handles a practical trade-off: do you want one steady, reproducible answer, or several different versions to choose from?
The one-sentence definition
Temperature is a number that decides how adventurous a model gets when it picks each next word. A model writes one word at a time, and at every step it holds a ranked list of candidate words by probability. Low temperature means it sticks to the top-ranked word, so the output trends toward deterministic and repeats almost identically; high temperature means it treats the lower-ranked candidates more generously, so the wording spreads out. Push temperature to its floor and you basically get a fixed, reproducible answer.
A concrete example
The clearest example is asking an AI to write the same product tagline several times: at low temperature the versions come out nearly identical, and at high temperature they diverge. Hitting “regenerate” at a low setting gives you conservative, near-identical phrasing; raise the temperature and those runs split into playful and formal takes, at the cost of an occasional clunky or off-topic line. You can feel this directly in a tool like ChatGPT, where generating the same prompt twice rarely returns the exact same text, and a randomness setting like temperature is what sits underneath. One caution: more variety does not mean more reliability. The model can be confidently wrong, so important claims and citations still need a human review.
How it differs from related settings
Temperature often gets mentioned alongside “random seed” and “top-p,” but they govern different things. A random seed fixes which random path a run takes, so pinning the seed reproduces the same result; temperature sets how wide that randomness is allowed to be. Top-p is a separate way to limit the candidate pool: it trims the long tail of unlikely words by cumulative probability, while temperature rescales the whole ranking. You can turn both knobs at once, but each one adjusts something distinct.
When it actually matters
Turn temperature down when you need steady, reproducible answers, and turn it up when you want a spread of different ideas. A few common cases:
- Field extraction, code, or fixed-format output: low temperature is easier to trust and easier to check.
- Naming, brainstorming, or several copy variants in one go: high temperature earns its keep, since you’d rather sift through options than get the same line twice.
Two things to keep in mind. First, even at a low temperature the model can still be wrong, so claims and numbers remain yours to verify. Second, whether you can set temperature at all, and how far, varies by tool, plan, and region, so treat the product you’re actually using as the source of truth.
