Hallucinations are the LLM's most damaging failure mode. The model doesn't 'know' things — it predicts statistically likely next tokens. When the prompt is in a domain the model has shallow training on, the prediction can be wrong while still sounding fluent and authoritative. Famous examples: lawyers citing AI-invented court cases (sanctioned in 2023), researchers citing AI-fabricated papers, support bots quoting non-existent product features. Mitigations: RAG (force the model to ground answers in real documents), explicit citations (require URLs the model can verify), 'I don't know' tuning (training the model to express uncertainty), and human review on anything consequential. The 2026 frontier models hallucinate dramatically less than 2023's, but they still hallucinate — never deploy an LLM in a critical workflow without a verification layer.
قاموس
ما هو Hallucination؟
When an LLM confidently produces information that's factually wrong — invented citations, fake quotes, plausible-sounding but incorrect claims.
مصطلحات ذات صلة
RAG (Retrieval-Augmented Generation)
A technique that lets an LLM answer questions about information it wasn't trained on — by retrieving relevant documents at query time and stuffing them into the prompt.
LLM (Large Language Model)
An AI system trained on massive text datasets to predict and generate human-like text — the technology behind ChatGPT, Claude, Gemini, and most modern AI chatbots.
RLHF (Reinforcement Learning from Human Feedback)
The technique that turns a base LLM into a useful assistant — by having humans rate model responses and using that feedback to fine-tune behavior.
العودة إلى قاموس AI