Hallucination

Simple Definition

Hallucination is when an AI model states something that is factually incorrect with complete confidence. It makes something up — a statistic, a name, a citation, a date — and presents it as fact.

The term comes from the idea of perceiving something that isn’t there.

Why It Happens

LLMs don’t access a fact database. They generate text by predicting what words should follow other words based on patterns in their training data. When they encounter a question they don’t have good training signal for, they may “fill in the gap” with plausible-sounding content rather than admitting uncertainty.

Common Types of Hallucinations

• Fake citations — making up paper titles, authors, or publication dates
• Wrong statistics — inventing numbers with false precision
• Nonexistent people or events — confidently describing things that never existed
• Incorrect facts — getting names, dates, or details wrong
• False legal or medical information — stating incorrect regulations or diagnoses

How to Protect Yourself

1. Verify before trusting — always check specific facts, statistics, names, and citations
2. Use sourced tools — Perplexity shows citations; you can check the sources
3. Ask the AI to be uncertain — prompts like “only state things you’re confident about” can help
4. Use RAG systems — grounding the AI in real documents reduces hallucinations
5. Be more skeptical of specifics — general explanations tend to be more reliable than specific claims

Related Terms

• LLM — the type of model most prone to hallucination
• RAG — a technique for reducing hallucinations by grounding in real documents
• Grounding — connecting AI responses to factual sources
• AI Safety — the broader field concerned with making AI reliable and trustworthy