Why AI gives confident wrong answers: a beginner's guide to hallucinations — companion videos

IBM Technology

Martin Keen sorts hallucinations into four named buckets — sentence contradictions, prompt contradictions, factual errors, nonsense — and walks through each on a lightboard. After the article gives you the why, this video gives you a vocabulary for spotting the type of mistake in the wild so you can decide how much to trust a given answer.

What you should get from this: You can name the four kinds of hallucination and recognize which one you are looking at when a chatbot gets something wrong.

Watch or know first: None — watchable cold.

Computerphile

A long, calm sit-down with AI-safety researcher Rob Miles on why a model that's trained to predict plausible text will always, structurally, sometimes invent facts. Recorded in early 2023, but the underlying mechanism it explains hasn't changed and the framing is still the cleanest way to internalise "it's a fluency engine, not a truth engine."

What you should get from this: You can explain why a model trained to predict plausible text will sometimes invent facts, no matter how polished it sounds.

Watch or know first: None, though it helps to have seen a short hallucination explainer like the primary pick first.