Temperature
Updated June 10, 2026
The dial that controls how adventurous a model's word choices are — and one reason default AI text is so predictable.
Definition
When a model picks its next token, it has a probability for every candidate. Temperature reshapes those probabilities before sampling: near 0, the model almost always takes the single most likely token (deterministic, safe, repetitive); around 1, it samples proportionally (more varied); higher, it takes real risks (creative shading into incoherent). Chat products typically run moderate temperatures tuned for reliability.
Temperature and detection
Low-temperature text is maximally predictable — the lowest-perplexity, most detectable output a model can produce. Higher temperature raises surface variety but doesn't restore human rhythm or substance: you get unusual words in machine-shaped sentences. That's why "just raise the temperature" isn't a humanizing strategy — structure, not vocabulary spice, is what detectors measure and what a real humanize pass changes.