AI Hallucinations

KEY TAKEAWAY

AI hallucination is when an AI generates false information confidently — as if it were true.

💡 EVERYDAY ANALOGY — BEFORE THE TECH

😴

Hallucination is like a confident friend who makes things up

Imagine asking a friend for a restaurant recommendation. They confidently give you a name, address and description. You get there and the place doesn't exist. Your friend wasn't lying — they genuinely "remembered" it. They constructed a plausible-sounding memory that felt real. AI does the same: it constructs plausible-sounding text, even when the facts are wrong.

👁️ Visual Explanation

AI generating a fake citation: Prompt → LLM generates plausible-looking text → Outputs a real-sounding but non-existent academic paper

🌍 In Plain English

AI models predict the next plausible word. If a made-up "fact" sounds statistically plausible based on its training, the AI will say it confidently. It has no ability to check if what it's saying is true.

Real-World Example

Lawyers have submitted AI-written legal briefs that cited court cases that don't exist. The AI generated real-looking case names, judges and rulings — all fabricated.

🎓 For Students

If you ask an AI for academic references, always verify them. The paper, journal and author it cites may not exist.

💼 For Professionals

Before using AI-generated data, statistics or quotes in a presentation, verify each one independently. A single confident hallucination can embarrass you in front of clients.

🔧 Technical DepthOptional — expand when ready

Hallucination occurs because LLMs optimise for plausible next-token prediction, not factual accuracy. Training on text teaches it "what sounds right" statistically, not "what is true". Techniques to reduce hallucination include retrieval-augmented generation (RAG), grounding outputs in verified sources, and training with human feedback (RLHF) that penalises false claims.

Key Terms

HallucinationWhen AI generates confident, plausible-sounding but false information

GroundingConnecting AI output to verified sources to reduce fabrication

RAGRetrieval-Augmented Generation — giving the AI access to real, current documents before answering

⚠️ Common Misconception

AI is not "lying". It has no concept of truth or falsehood. It generates what statistically follows from the context — and sometimes that produces false information.