Hallucinations
When AI answers with total confidence… and is making it up.
The analogy
We all have that friend who never says “I don't know”: ask them something they have no clue about, and they'll improvise a convincing answer without blinking. Models do the same — they're designed to continue text plausibly, not to stay quiet when unsure.
In detail
A hallucination is output that is false but plausible: the model generates the most likely text according to its patterns, without checking facts against any source. It happens most with specific data (dates, figures, citations, URLs) and topics underrepresented in training. Mitigations include RAG, instructions like “say so if you're not sure”, and verifying anything important.
An example
You ask for five scientific papers on a topic and the model returns perfectly formatted titles, authors and years… for papers that don't exist. Always verify references.