Hallucination

Language models generate statistically plausible continuations, and plausible is not the same as true: invented citations, wrong dates, and nonexistent API functions can read exactly like correct output. Hallucination rates have fallen with better training and calibration — newer models more often say they are unsure — but the failure mode is inherent to generative prediction and has not disappeared.

The standard mitigations are grounding and verification: retrieval-augmented generation so claims rest on provided sources, citation requirements that make answers checkable, structured outputs validated against schemas, execution checks for generated code, and human review where errors are costly.

Mitigation costs tokens — retrieved context, verification passes, judge models — creating an explicit reliability budget. Spending a few extra cents per request on grounding is usually cheap insurance against the business cost of a confidently wrong answer.