Hallucination

A hallucination is when an LLM generates confident-sounding information that is false, fabricated, or unsupported. It invents facts, citations, names, dates, and quotes that never existed.

In one sentence

A hallucination is when an AI makes something up and says it confidently, like a student bluffing an exam answer.

What it's for

Hallucinations are the #1 risk when using LLMs for research, journalism, law, and medicine. They sound authoritative but are completely fabricated. Every major LLM hallucinates; none have solved the problem.

How to think about it

Think of an LLM as a very good improviser. Asked a question, it invents a plausible-sounding answer rather than admitting ignorance. It doesn’t know it’s lying—it simply predicts the most likely next words, and sometimes that means inventing facts.

Common misunderstandings

Users think newer models hallucinate less. They do—slightly—but the problem is not solved. GPT-4, Claude 3, Gemini Ultra—all hallucinate. The only protection is verification: always check facts against authoritative sources.

If you want to try it

Ask ChatGPT: “Give me 3 scientific studies linking coffee to longer lifespan, with authors and publication years.” Then search PubMed for those studies. See how many are real.