Hallucination (Hallucination)

## Why It Is Called "Hallucination" Just as human hallucinations involve perceiving things that do not exist, AI hallucinations also generate "facts that do not exist." However, unlike human hallucinations in a fundamental sense, LLMs have no mechanism for determining whether something is factual. They simply generate the next most probable token in a chain, and the result may happen to align with reality or may be entirely fabricated. ## Typical Patterns Hallucinations take several characteristic forms. Representative examples include citations of non-existent papers (generating fictitious author names and DOIs), false biographical details attributed to real individuals, and the fabrication of plausible-sounding numerical data. What makes this particularly troublesome is that hallucinated output is grammatically correct and blends naturally into context. Obvious errors are easy to spot, but the pattern of being "90% correct and 10% false" is what makes detection so difficult. ## Mitigation Approaches The most promising mitigation approach at present is the adoption of RAG (Retrieval-Augmented Generation). Before the model generates a response, it retrieves relevant information from an external knowledge base and uses that information as the basis for its answer, thereby increasing the probability of factually grounded output. Another direction is the incorporation of HITL (Human-in-the-Loop). By designing a workflow in which humans review AI output, the risk of hallucinations making their way into final deliverables is reduced. In fields such as medicine and law, where the cost of misinformation is high, this combination is becoming the de facto standard.