Hallucination as Matrix-Crossing

In the AI Story

The framing reveals a tradeoff that the engineering community has not fully confronted. Techniques that reduce hallucination also reduce the probability of genuine bisociation. Retrieval-augmented generation, grounding mechanisms, and tighter output constraints all work by keeping the machine more firmly within the matrix specified by the prompt. They increase accuracy by decreasing divergence. And decreased divergence means decreased matrix-crossing, which means a reduced probability that the output will introduce elements from an unexpected domain that reveal a structural identity the user had not perceived.

This does not imply that hallucination should be tolerated in applications where accuracy is paramount—legal research, medical diagnosis, financial analysis. In these contexts, the engineering effort to reduce hallucination is entirely justified. The implication is that the creative use of the machine and the reliable use of the machine pull in opposite directions along the temperature continuum, and the practitioner must navigate the tension consciously.

The kinship also clarifies why pseudo-bisociation is such a dominant AI failure mode. A pseudo-bisociation is a hallucination at the structural level rather than the factual level: the machine produces a connection that appears to reveal structural identity but actually exploits surface resemblance. The machine cannot distinguish between matrix-crossings that find something and matrix-crossings that find nothing, because it has no independent access to whether structural identities actually hold. That determination requires a prepared human frame.

The practical consequence is that the creative use of AI requires a specific tolerance for hallucination-adjacent behavior. The practitioner who insists on zero hallucination will also minimize bisociation. The practitioner who accepts high hallucination will drown in noise. The productive zone is the middle—where the machine is permitted enough divergence to produce genuine cross-matrix connections, and the human is disciplined enough to verify which crossings find structural identity and which find nothing.

Origin

The term 'hallucination' entered AI vocabulary in the 2010s, originally applied to computer vision systems that produced confident identifications of absent features. The extension to language models became standard after 2020, and the structural kinship with creative frame-crossing became visible as the same systems began to be used for creative rather than purely informational tasks.

Key Ideas

Same operation, different outcomes. Hallucination and bisociation both cross matrix boundaries; the difference is whether the crossing finds structural identity.

Engineering tradeoff. Reducing hallucination reduces bisociation; the two cannot be simultaneously maximized at the architecture level.

Pseudo-bisociation as structural hallucination. The dominant creative failure mode is hallucination not of facts but of structural connections.

Context-dependent evaluation. Accuracy-critical applications should minimize divergence; creative applications require tolerating it.

Verification as human responsibility. The machine cannot distinguish productive crossings from spurious ones; only a prepared human frame can.