Temperature
Temperature is a setting that controls how much randomness a language model uses when choosing each next token — low values make it pick the most likely option almost every time, high values let it reach further down the list.
Last reviewed 15 August 2026 · The HumanFlow team
In plain English
At temperature 0 the model takes its top choice at every step, so the same prompt gives you nearly the same answer every time.
Turn it up and less likely words start getting picked. The text becomes more varied, and eventually less coherent.
A worked example
The same opening continued at three settings.
Prompt: "The most surprising thing about the results was" temp 0.0 → that they were consistent across all three cohorts. temp 0.7 → how little the second cohort moved, given its size. temp 1.4 → the silence afterwards, which nobody had budgeted for.
The first is the safest continuation available and reads like every other sentence of its kind.
The second is still plausible but has picked a specific detail rather than the generic one.
The third is genuinely unexpected and is also where incoherence starts creeping in at longer lengths — high temperature buys surprise at the cost of control.
Why it matters for AI detection
It explains a pattern people notice without being able to name: default settings on most consumer tools sit low, which is why generated text so often reads as flat, safe and interchangeable.
It is directly connected to detection. Low-temperature output is by construction the text a model found most probable, which is exactly what a perplexity-based detector scores as machine-like. Higher temperature raises perplexity.
That is not a recommendation. Turning temperature up to change a score is the bypass framing this site does not offer, it makes output less accurate rather than more yours, and no setting changes whether you are required to disclose what you used.
Commonly confused with
- Perplexity
- Temperature is a knob on the generator; perplexity is a measurement made afterwards by a different model. They move together but one is a setting and the other is an observation.
- Top-p / nucleus sampling
- Another way of controlling the same randomness, by restricting choices to the smallest set covering a probability threshold rather than by reweighting all of them. Tools often expose both.
Read next
Part of the AI detection glossary.