Hallucination-As-Feature

The Epistemological Premise

A hallucination, in the AI sense, is a confident error — the system produces a specific false detail as if it were true. Standard practice is to identify and remove all hallucinations, on the grounds that factual accuracy is the baseline requirement for trustworthy output. The Foundry does not dispute this for information retrieval contexts.

Hallucination-As-Feature is a narrower claim: that in creative and conceptual contexts, a specific class of AI error can reveal a structural truth that would not have been found through accurate generation. The error is not an accidental approximation of a fact. It is an accidental approximation of a shape — a metaphor, a compression, a frame — that happens to be more accurate to the phenomenology of a situation than the correct fact would be.

The Test

How do you distinguish a productive hallucination from a simply wrong one? The Foundry proposes three diagnostic questions:

• Is it more true than accurate? Does the invented detail capture something real about the texture or meaning of the situation, even if it is factually incorrect?
• Can it be sourced to something real? Can the hallucinated detail be reverse-engineered to a genuine reference — a real event, a pattern, a known phenomenon — that the model was approximating?
• Does it open productive territory? Does the invented detail generate an interesting question, a new frame, or a conceptual connection that the correct fact would have foreclosed?

If all three answers are yes, the hallucination is a candidate for Hallucination-As-Feature. The practitioner preserves it, labels it as invented, and uses it as a starting point rather than a conclusion.

Usage in context: "The AI called it 'violently beige.' That's technically wrong. But it's exactly right. Keep it — it's a feature."