Outcome. You can classify a factual failure, choose an appropriate grounding or verification control, and define when the system should abstain or escalate.
A hallucination is generated content that is unsupported, inconsistent with the provided source, or false relative to the world. It is not caused by one simple flaw. Training objectives reward likely continuations, data contains gaps and errors, prompts can ask for unknowable specifics, retrieval can fail, and evaluations often reward answering more than admitting uncertainty. Tool-equipped systems can check external sources, but only if the application asks, the tool works, and the result is used correctly.
Separate failure types. An intrinsic error contradicts supplied evidence; an extrinsic claim adds unsupported information; a retrieval failure provides the wrong evidence; a citation failure attaches a real source to a claim it does not support. Each needs a different fix.
Confidence must be calibrated over many cases: among answers labeled 80% confident, roughly 80% should be correct under the defined conditions. Fluent language and a self-reported percentage are not calibration. For high-risk tasks, require sources, deterministic checks, abstention on missing evidence, and qualified human review.
Mental model. Reliability is a pipeline: retrieve or observe evidence, generate a claim, verify support, calibrate uncertainty, then decide whether to answer or escalate.
Evidence trail — reviewed 23 July 2026. TruthfulQA studies imitation of common falsehoods: https://arxiv.org/abs/2109.07958. Calibration research: https://arxiv.org/abs/2012.00955 and https://arxiv.org/abs/2207.05221. Treat old benchmark numbers as historical, not current model rates.
Require claims as {claim, evidence_quote, source_id, support: entailed|partial|none}. Reject or abstain when the quote does not entail the claim. Test missing policy, conflicting versions, irrelevant retrieval, and a malicious document. Route policy exceptions to a human.