Temperature

Temperature is a setting that controls how random or predictable an AI language model's output is. When a model generates text, it is repeatedly choosing the next token from a range of possibilities, each with its own probability. Temperature adjusts how boldly it strays from the most likely choice. A low temperature makes the model conservative, favoring the safest, highest-probability options, so its output is focused, consistent, and predictable. A high temperature loosens things up, giving less likely words a better chance of being picked, which makes the output more varied, surprising, and creative, but also more prone to wandering or error.

The mechanics work on the probability distribution the model produces at each step. For every possible next token, the model assigns a likelihood. Temperature reshapes that distribution before a choice is made. At a low value, the differences between options are sharpened, so the top choice dominates and the model almost always picks it. At a high value, the differences are flattened, so lower-ranked options become competitive and the selection gets more adventurous. A temperature near zero makes the model nearly deterministic, tending to give the same answer to the same prompt every time, which is useful when you want reliability. Higher values introduce genuine variety, so the same prompt can yield different responses on different runs. There is no single correct setting; the right value depends on the task.

The term comes from the Latin temperatura, meaning a proper mixing or proportion, and it borrows an idea from physics, where higher heat means more energetic, disorderly motion. In AI the metaphor holds: higher temperature means more disorder and unpredictability in the output, lower temperature means more order and restraint. The name signals that you are tuning how energetic the model's word choices are.

For a business using AI tools, temperature matters because it lets you match the model's behavior to the job. For tasks that demand accuracy and consistency, such as extracting data, answering factual questions, generating code, or producing standardized responses, a low temperature keeps the model on the rails and reduces the chance of odd or invented output. For tasks that benefit from variety and fresh ideas, such as brainstorming campaign concepts, drafting several different headlines, or writing creative copy, a higher temperature produces a wider range to choose from. Knowing this setting exists gives you a practical lever over the tradeoff between reliability and creativity.

The nuances and common mistakes are important. A high temperature does not make a model smarter or more insightful; it just makes it less predictable, and pushed too far it produces incoherent or inaccurate text. A very low temperature can feel repetitive and flat, giving nearly identical phrasing every time. Temperature is also not a truth dial: lowering it reduces randomness but does not guarantee correctness, since a model can be confidently wrong at any setting. It works alongside other controls that shape output and connects to the broader behavior of generative AI and the foundation models it runs on, to the context window that determines what information the model is working from, and to prompting techniques like zero-shot prompting, where clear instructions and a sensible temperature together shape the result.