In AI, a hallucination happens when a model produces content that sounds convincing but is false, unsupported, or not grounded in the information it was given. The term is used most often with language models, but similar failures appear in image, audio, and multimodal systems as well. Hallucinations matter because they are easy to miss. A polished answer can still be wrong.
Why Hallucinations Happen
A language model is trained to continue patterns, not to guarantee truth. If the prompt is ambiguous, the supporting material is weak, or the model is pushed beyond what it knows, it may fill in gaps with likely-sounding output. The model is optimizing for plausible continuation, not independent fact-checking.
Hallucinations can show up as invented citations, incorrect summaries, fabricated numbers, broken instructions, or overconfident answers to questions that should have been refused or qualified. They are especially risky in medicine, law, finance, research, and any workflow where specific details matter.
How Systems Reduce Hallucination
No single technique eliminates hallucinations, but several can reduce them. Better prompts help. Grounding the model in trusted sources helps more. RAG can bring in relevant evidence at answer time. Guardrails, tool-based verification, and human review are also important in higher-stakes systems.
The key lesson is that fluency is not proof. AI output should be treated as a useful draft or reasoning aid unless the workflow has explicit checks for truth, source quality, and permission boundaries.
Related concepts: Grounding, RAG, Guardrails, Context Window, and Large Language Model (LLM).