"Just lower the temperature" is the most common advice for reducing hallucination. It is also mostly wrong. Temperature controls how random the model's token selection is. It does not control whether the selected tokens are true. Low temperature means the model picks the most likely output - but the most likely output can still be wrong.
Temperature is a parameter that adjusts the probability distribution over the model's next-token predictions. At temperature 1.0, the model samples normally from its predicted distribution. At temperature 0.0, it always picks the single most likely next token (greedy decoding).
High temperature (0.7-1.0): More random, more creative, more varied output. The model explores lower-probability tokens, producing novel combinations.
Low temperature (0.0-0.3): More deterministic, more repetitive, more predictable output. The model sticks to high-probability tokens, producing the "safest" continuation.
Notice what is missing from this description: accuracy. Temperature controls the spread of the distribution. It does not change the distribution itself. If the most likely token is wrong, temperature 0 will reliably produce that wrong token every single time.
Temperature 0 does not mean "always correct." It means "always picks the same answer." If that answer is wrong, you get the wrong answer with 100% consistency.
Consistent hallucinations. At low temperature, the model produces the same output every time for the same input. If that output contains a hallucination, the hallucination is perfectly reproducible. Self-consistency checks - asking the model the same question multiple times and comparing answers - will show agreement. The hallucination passes the consistency test because it is deterministically wrong.
Harder to detect. Variance in output is actually a useful signal for uncertainty. When the model gives different answers at high temperature, that tells you it is uncertain. At low temperature, the model gives one confident answer even when it should be uncertain. You lose the uncertainty signal.
No creative recovery. At higher temperatures, the model sometimes "stumbles" onto the right answer through a less likely reasoning path. Low temperature eliminates this. The model is locked into its most likely path, which may be the wrong one.
Low temperature is useful for tasks where the model already knows the correct answer and you want consistent formatting. Code generation with well-known patterns, structured data extraction, simple factual lookups from training data.
It also reduces creative hallucinations - cases where the model invents novel but false details through random token selection. A model at temperature 1.0 might hallucinate a creative but wrong package name. At temperature 0, it will hallucinate the most common wrong package name instead. The hallucination is still there. It is just more predictable.
Temperature is a style knob, not an accuracy knob. It controls how the model selects from its predictions. It does not improve the predictions themselves.
If the model's training data is wrong, temperature 0 gives you the wrong answer with full confidence. If the model's training data is outdated, temperature 0 gives you the outdated answer with full confidence. If the model's knowledge has gaps, temperature 0 fills those gaps with the most likely (but still fabricated) content.
The fix for all three cases is the same: give the model access to reality. A verification layer that injects real-time, live information into the model's context before it generates output, and validates the output before it reaches the user.
Check does not adjust the model's temperature. It gives the model the actual facts. When the model has verified data about the real state of the environment, it does not need to guess - at any temperature. The answer comes from reality, not from probability.
120 verifications a day free. No card, no signup.