AI Temperature

The plain-English definition

Temperature is a numerical setting that controls how "creative" an AI model's outputs are. At temperature 0, the model picks the most likely next word every time — predictable, factual, consistent. At temperature 1.0+, it samples from a wider distribution — more varied, more creative, sometimes weirder.

The mental model

Imagine asking 100 people to finish the sentence "The sky is ___." Most would say "blue." That's the predictable answer (low temperature). A creative writer might say "the color of bruises healing" — accurate, but non-default. That's high temperature.

When to use which

• Temperature 0.0–0.2: Factual Q&A, code generation, classification, structured outputs. You want consistency.
• Temperature 0.3–0.7: Standard writing, summaries, helpful answers. The default for most models is around 0.7.
• Temperature 0.8–1.0: Brainstorming, creative writing, generating variations.
• Temperature 1.0+: Wild brainstorming, poetry, unusual associations. Can produce gibberish.

Where to find it

Most consumer chat interfaces (ChatGPT, Claude, Gemini) hide the temperature slider — you get a sensible default. To control it, you need:

• The API of any major provider
• OpenAI's Playground (free)
• Tools like LM Studio or local LLM apps

Practical tip

If your AI keeps giving generic answers, try asking for "10 wildly different options" — this approximates higher temperature in a regular chat. If you want consistency, ask for "the most likely answer" or "stick to facts in [SOURCE]."

Related terms

• Hallucination
• System Prompt
• Token