Hallucinations

In everyday AI talk, a hallucination is when a model produces confident content that is not grounded in reality—fake citations, invented biographies, wrong numbers—while sounding fluent.

Fluent falsehoods are more dangerous than awkward uncertainty. Overtrust is how hallucinations become real-world harm (Course 29 medical overtrust; scams that invent “official” stories).

The model is rewarded for producing likely token sequences, not for checking a world model. If a plausible-sounding answer is probable under training patterns, it may emit it even when false.

Mitigations (reduce, not eliminate): retrieval (RAG), tools/calculators, ask for sources and then verify them yourself, constrain to provided documents, human review on high stakes.

A chatbot invents a court case name with a realistic ID format. It looks official. It is not.

Ask for a source, then open the link or search the citation. If it does not exist, you just caught a hallucination.

⚠️ Myth: Hallucinations mean the model is “lying.”

✓ Reality: It is generating plausible tokens without a guarantee of truth.

⚠️ Myth: Temperature 0 removes hallucinations.

✓ Reality: It reduces randomness; it does not install a knowledge verifier.

• Holistic evaluation / factuality literature overview via Hugging Face LLM course discussions: https://huggingface.co/learn/llm-course/ ↗
• NIST AI RMF (trustworthiness characteristics): https://www.nist.gov/itl/ai-risk-management-framework ↗
• ANN live: https://www.ainerdnetwork.com/learn/hallucinations ↗