Plausible-sounding model outputs that are wrong or unsupported—especially risky for facts, APIs, and security assumptions.
Hallucination is not random noise: models optimize for fluent continuation, so they can invent citations, parameters, or behaviors that look right. Mitigations include retrieval (RAG), verification, tests, human review, and structured outputs where appropriate.
Study angle
Treat hallucination as a default failure mode, not an edge case—similar to how you assume network timeouts in distributed systems.