Safety and governance
Hallucination
What is hallucination?
A hallucination is model output that is fluent, confident and false. It arises because language models are trained to produce probable text rather than true text, and probability and truth diverge whenever the model lacks the knowledge.
Fine-tuning can reduce it by teaching the model to decline, but only if refusals appear in the training data. A dataset containing exclusively confident answers teaches the model that a confident answer is always expected.
Retrieval reduces it by supplying the facts. Neither approach eliminates it, and systems where a confident error is expensive need verification rather than only mitigation.
Related terms
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.