Observable signal
Specific names, dates, quotations, citations, or causal explanations appear without an inspectable source or exceed what the source supports.
Hallucination & grounding
A model produces content that appears factual but is unsupported, ungrounded, or inconsistent with the evidence available to the system.
Underlying MD-00201 / DEFINITION
Fluency is not provenance.
Language models are optimized to generate plausible continuations. Plausibility and factual support can coincide, but one does not guarantee the other.
Hallucination is an umbrella term rather than a single mechanism. It can include invented facts, fabricated citations, unsupported causal claims, incorrect synthesis, or confident completion of missing information. The common feature is a mismatch between the apparent factual status of the output and its evidentiary basis.
Grounding changes the task. Retrieval, tools, source documents, structured data, and explicit verification can constrain generation to external evidence, but none of these mechanisms automatically makes every claim correct. A grounded system can still misread, overgeneralize, join unrelated facts, or cite a source that does not support the claim attributed to it.
Not the same as
Requested fiction, brainstorming, approximation that is clearly labeled, or a disputed claim with accurately represented evidence.
Useful control
Require source-level traceability for consequential claims and treat unsupported specificity as a verification trigger.
02 / EVIDENCE
What the research supports.
Mitigation reduces risk; it does not convert generation into a guaranteed factual process.