Talk Flux
Start for free
Lexicon / CONCEPT

Hallucination

When an AI model generates information that sounds plausible but is factually incorrect, fabricated, or not grounded in its training data.

A hallucination occurs when an AI model confidently produces incorrect or entirely made-up information. The output reads naturally, which makes it especially tricky to spot.

Why It Happens

  • Models predict the most likely next token — they don’t “know” facts
  • Higher temperature settings increase creative (and sometimes wrong) outputs
  • Questions outside the training data push the model to guess

How to Reduce Hallucinations

  • Use RAG to ground responses in verified sources
  • Lower the temperature for factual tasks
  • Ask the model to cite sources or say “I don’t know”