AI Hallucinations

The learner wants to understand AI hallucinations. It is unclear whether they are approaching this from a practical user perspective, a technical ML perspective, or a safety/policy angle, and how deeply they already understand how language models work. What Is a Hallucination: Ask the learner to distinguish a hallucination from a refusal and from a factual error caused by outdated training data. Why They Happen: Ask the learner to explain in plain terms why a next-token predictor can produce a confident falsehood even with no intent to deceive.

The gap most people have on ai hallucinations is the part that actually changes outcomes: The learner wants to understand AI hallucinations. Once that lands, the supporting ideas — detecting hallucinations — start paying off in everyday decisions.

Many people first hear "hallucination" and think of the ai is generating false or made-up information presented as if it were true. That framing is exactly right: a hallucination is a fluent, confident output that has no grounding in fact, evidence, or the provided context. Many people first hear "confabulation" and think of confabulation means the model fills gaps with plausible-sounding invented detail, borrowing the neuropsychology term for gap-filling memory errors. The term emphasises that the model, like a confabulating patient, is not lying but genuinely producing what its internal process treats as a coherent completion — making the mechanism clearer than the vaguer word hallucination.

LearnBench teaches ai hallucinations in 6 adaptive cards organized around 4 core ideas. A few quick checks find what you already know, then the lesson skips it — so you only see the parts you're actually missing, framed with concrete analogies.