What is Hallucination (AI)?
Turkish: Halüsinasyon (LLM Uydurması)
AI hallucination is when a model produces information that sounds plausible but is false, unsupported, or not grounded in the source.
What is Hallucination (AI)?
Hallucination is when a model produces fluent and confident text that is not backed by a reliable source. A wrong date, invented regulation clause, nonexistent product code, or faulty technical recommendation can all be hallucinations.
This happens because LLM models generate text probabilistically. A model does not always choose to say “I don’t know”; with missing context, outdated training data, or an ambiguous question, it may create a sentence that sounds reasonable.
Why It Happens
Hallucination risk usually increases in three cases: the model lacks access to the right source, the user question is ambiguous, or the system does not verify the answer against evidence. Long contexts can also cause important details to be overlooked.
RAG reduces the risk by grounding answers in company documents or current data. It does not remove the risk entirely; if the wrong document is retrieved or the model moves beyond the evidence, hallucination can still occur.
Protection and Business Risk
In customer support, legal, finance, healthcare, integration, and operations workflows, invented information can cause direct harm. Critical systems should use citations, scoped answers, human approval for low-confidence cases, and AI guardrails.
Hallucination is not a bug that can be fully eliminated. It is a risk to measure and manage through test sets, error examples, and user feedback.
Related Terms
AI guardrails are control layers that constrain model inputs, outputs, and tool use against safety, policy, and quality rules.
LLM (Large Language Model)An LLM is a model trained on large text datasets that can understand and generate natural language, forming the basis of tools like ChatGPT.
RAG (Retrieval-Augmented Generation)RAG is an AI architecture where a language model retrieves relevant passages from documents or databases before generating an answer.