Concept

Hallucination (in LLMs)

A hallucination is what we call a language-model output that sounds confident, looks right, and is wrong. The model invented a study that doesn't exist. It cited a court case that was never filed. It returned a function signature that compiles in your head but not in any language. The output is not deceptive in intent (the model has no intent); it is the natural failure mode of a system that predicts the next word from patterns rather than from grounded facts.
Reviewed by 7wData

Why it matters

Hallucinations are the reason “generative AI” is not the same product as “answer engine.” A chatbot that hallucinates a refund policy is embarrassing. An agent that hallucinates a transaction reference is a liability event. Every serious LLM deployment in 2026 wraps the model in some form of grounding (retrieval, citations, evidence trails) precisely because the raw model output is not trustworthy by default. The harness around the model is what makes the model usable; without it, you ship the failure mode straight to the user.

Where you’ll encounter it

You will encounter hallucinations in three places: in unsourced chatbot answers (the most visible kind), in code completions that look plausible but reference nonexistent APIs, and in summarization tools that introduce facts the source document did not contain. The third one is the most dangerous because it is the hardest to catch: the reader trusts the summary precisely because the source is too long to verify against. Reduction patterns include retrieval-augmented generation (RAG), forced citation, evaluation gates, and human review on consequential outputs.

Part of the 7wData AI Glossary. Tracking how concepts like this move in the expert conversation: daily signals at ins7ghts.com.