Temperature
Also: Sampling temperature, Temperature setting, Temperature parameter, Temperature (LLM), Parametre de temperature, Temperatur-Parameter
A setting that controls how random or predictable an AI model's output is: low keeps it safe and consistent, high makes it more varied and creative.
What It Is
Temperature is a setting that controls how much randomness a generative AI model uses when it produces text. At each step the model has a ranked list of possible next words with probabilities attached. Temperature decides whether the model almost always picks the single most likely word or spreads its choices across less obvious options. A low temperature (near 0) makes output focused and repetitive. A high temperature (above 1) makes it more varied and surprising, sometimes at the cost of accuracy.
Why it matters
Temperature is one of the few AI dials a business team can adjust without retraining a model, and it directly shapes brand risk and output quality. A support chatbot answering compliance questions needs a low temperature so it gives the same reliable answer every time, while a campaign brainstorming tool benefits from a higher temperature to surface fresh angles. Setting it wrong is expensive in either direction: too low and your content feels robotic and identical across customers, too high and the model drifts into off-brand or factually loose territory. Leaders reviewing an AI vendor or an internal deployment should ask what temperature is configured and why, because it reveals whether the team optimized for consistency or creativity.
How it works
In practice temperature is exposed as a numeric field in an API call or a settings panel, often between 0 and 2. A marketing lead running a subject-line generator might test the same prompt at 0.2, 0.7 and 1.2 to see where variety peaks before quality drops. The value works alongside the prompt and other controls: a tight prompt with low temperature produces near-deterministic results useful for legal or financial text, while a looser prompt with moderate temperature suits ideation. It does not change what the model knows, only how it selects from what it already predicts. A common working pattern is to keep temperature low for anything customer-facing and factual, raise it for internal ideation, and document the chosen value so results can be reproduced and audited later.
Frequently asked questions
What does temperature do in a generative AI model?
Temperature controls how much randomness a model uses when it picks each next word. At every step the model holds a ranked list of possible words with probabilities; a value near 0 makes it almost always take the most likely one, while a value above 1 spreads the choice across less obvious options. It does not change what the model knows, only how it selects from what it already predicts.
What is the difference between a low and a high temperature setting?
Low temperature produces focused, consistent and sometimes repetitive output; high temperature produces varied and surprising output, sometimes at the cost of accuracy. The practical consequence is a trade-off between reproducibility and creativity. Too low and your content feels robotic and identical across customers, too high and the model drifts off-brand or becomes factually loose.
Why should a business leader care about a technical setting like temperature?
Because it is one of the few AI dials a business team can adjust without retraining a model, and it directly shapes brand risk and output quality. A support chatbot answering compliance questions needs a low temperature to give the same reliable answer every time; a campaign brainstorming tool benefits from a higher one. Asking a vendor or an internal team what temperature is configured and why reveals whether they optimized for consistency or for creativity.
Where do you actually set the temperature value, and what range is typical?
Temperature appears as a numeric field in an API call or in a settings panel, usually on a scale from 0 to 2. There is no universally correct value: it depends on the use case and on the prompt it works alongside. A tight prompt with low temperature gives near-deterministic results useful for legal or financial text, while a looser prompt with moderate temperature suits ideation.
How do I find the right temperature for a given use case?
Run the same prompt at several values and compare. A marketing lead building a subject-line generator might test at 0.2, 0.7 and 1.2 to see where variety peaks before quality drops. A common working pattern is to keep temperature low for anything customer-facing and factual, raise it for internal ideation, and document the chosen value so results can be reproduced and audited later.