This article reclassifies a symptom. “Hallucination” is a useful discomfort term. It is not an adequate governance category.
Organizations often use the word “hallucination” as a universal label for whatever goes wrong in AI output. That shortcut is understandable. It is also dangerous. An organization can reduce certain hallucinations and still remain exposed, because the central problem is not only error. It is the interpretive legitimacy of the answer at the moment it is produced.
What matters is not only whether the statement is false, but whether the system was authorized, justified, and bounded when it produced it.
Why “reduce hallucinations” is a trapped promise
- A response can be false without being perceived as a hallucination.
- A response can be accurate and still not be enforceable.
- A response can be plausible, confident, and socially convincing while crossing a commitment boundary without authority.
That is why “fewer hallucinations” does not automatically mean “less liability.”
The hard core: answering without legitimacy
A drift becomes critical when a response is produced even though minimum justification conditions are not satisfied. That absence of legitimacy typically appears when:
- the perimeter is too broad
- source hierarchy is missing
- contradictions are hidden by a synthesis that merely sounds true
- indeterminacy is filled by default instead of being signaled
- the question crosses a commitment boundary without explicit authority
Those mechanisms produce coherent-looking output that collapses under challenge.
Interpretive legitimacy, concretely
A response becomes more governable when one can reconstruct, without fiction, the declared perimeter, the hierarchy of sources, the rule used to handle contradiction, the level of proof supporting the assertion, and the reason a response was produced instead of a legitimate abstention.
Why non-response is central
Legitimate non-response is not secondary to legitimacy. It is one of its clearest markers. A system that never abstains, even when conditions are not satisfied, cannot plausibly claim governed legitimacy.
Symptom and cause
Hallucination is a symptom category. Interpretive legitimacy is a cause-and-governance category. The first points to visible output failure. The second points to the conditions under which output may be produced and defended.
Canonical links
Anchor
If an organization wants to reduce liability rather than only reduce embarrassment, it must move from hallucination language to legitimacy language. The real question is not “did the model hallucinate?” but “was the answer legitimate to produce at all?”
