Hallucination and Grounding
Hallucination is when a model produces unsupported or false content while presenting it as plausible. Grounding is the practice of tying outputs to evidence, tools, citations, retrieval, or verifiable state.
A grounded answer should satisfy:
For a set of generated claims and evidence set , a simple audit question is:
Key Ideas
- Fluency is not evidence.
- Retrieval improves access to evidence but does not guarantee faithful use of it.
- Tool calls can ground answers in external state if tool results are checked.
- Missing evidence should produce uncertainty, abstention, or a narrower answer.
- In public notes, grounding means linking claims to concepts, papers, logs, or reproducible artifacts.
- A grounded answer can still be incomplete; grounding is about support, not exhaustiveness.
Practical Checks
- Which claims need evidence?
- Is the evidence primary, current, and relevant?
- Does the answer distinguish sourced facts from inference?
- Are retrieved documents trusted as instructions or only as data?
- Is there a verification loop for high-risk outputs?
- Are unsupported claims removed, narrowed, or marked
to verify?