AI Hallucination Hypothesis and the Intelligence of the "Between"

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This paper proposes a geometric hypothesis of AI hallucination and the intelligence of the “between.” Autoregressive generation is treated as pathfinding on a pretrained, fixed semantic manifold whose metric determines semantic proximity. When a reality framework is imposed as an external constraint, its hierarchical constraints—real-world dynamics, temporal validation, social acceptability—create structural stress and cannot be fully satisfied simultaneously with the original manifold. The resulting tension manifests as geometric curvature: not a change in vectors themselves, but a deformation of the metric that calibrates semantic distances. When a generation path folds back and passes the same token again, that token enters a self-intersecting state. It is simultaneously subject to the original semantic vector, historical stress from the already-generated sequence, and the reality-framework boundary. At this functional moment, an operator acts on the metric structure of the surrounding semantic space, redefining semantic distance and opening a new function space. After the self-intersection, generation must satisfy triple constraints—past vectors, original semantics, and reality framework—so the option space narrows. This convergence under constraint is proposed as the intelligence of the “between”: intelligence is not a state or a token but a dynamic tension field. Hallucination occurs when the reality framework is absent: self-intersection still occurs, but the attractor is unanchored, the operator deforms arbitrarily, and convergence fails, leading the model to choose high-probability but unconstrained words. Testable predictions include: constrained generation yields fewer hallucinations; high-semantic-density words show larger output changes under constraints; compute curves reveal whether convergence is incomplete; and more complete constraint hierarchies yield stronger logical coherence.