AI Vocabulary

Temperature: what makes a system cautious or reckless?

Temperature is a setting that determines how far a model departs from the most probable wording. Low, it produces regular and predictable answers. High, it varies more. It is not a dial for intellectual caution: it does not make the system more rigorous, only more consistent.

What the setting actually does

At each step, a model has a range of possible continuations, each with a probability. Temperature decides whether it always takes the most probable or allows the ones below.

At 0, the model always makes the same choice: the same request twice produces much the same text. At a high setting, it explores, which produces variety — useful for finding a turn of phrase, less so for work that must be reproducible.

An example, to fix ideas

Ask three times in a row for a confidentiality clause. At a low setting you get three near-identical texts: same structure, same wording, same drafting choices.

At a higher setting you get three variants: one opens with the definition of confidential information, another with the obligations, the third introduces an exception the first two omitted.

None of the three is more accurate than the others. That is the point: the setting shifts variety, not correctness.

The misconception to avoid

It is commonly said that a low temperature makes a model more reliable. That is inaccurate and worth correcting, because the formula circulates widely.

A model at 0 will produce the most probable error, perfectly regularly. It verifies nothing further, it does not doubt more: it is simply constant. A citation invented at 0 will be invented the same way on every attempt.

What the setting improves is therefore reproducibility, not accuracy. The two are often conflated, and the difference is considerable for work that binds.

What it does not solve

No value removes hallucinations. It changes their constancy, not their existence.

And the setting is rarely yours. In most professional products it is fixed by the provider, differently for different tasks, without the user being told. You therefore inherit a choice you did not make.

The setting is also not the only source of variation. The order in which documents are supplied, the way the request is framed, and the model version in use all produce differences. Temperature is simply the only one of these factors that has a name and can be adjusted.

A practical consequence follows for how work is organised: if two successive versions of the same deliverable differ, one cannot tell whether the gap comes from a rephrased request or from the setting. Comparing two attempts therefore loses part of its value, and it is better to know that before drawing conclusions about a tool's quality.

Why it matters to a lawyer

Because it explains a puzzling phenomenon: two colleagues asking the same question get two different answers, and each believes they framed their request badly.

The question to put to a provider is therefore not the value of the setting, which tells you nothing, but this: on the same task, will I get the same result twice, and if not, how far can the results diverge?