A Relational Model of Human–AI Hybrid Intelligence

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).

Human–AI collaboration is often described as augmentation or division of labour. This paper proposes a narrower and testable account of how a sustained coupling may become epistemically productive: generative AI performs combinatorial expansion, while a human investigator may perform structural compression by identifying invariants, version conflicts, and candidate causal organization across a long project history. The theory distinguishes relationally coupled hybrid inquiry (a reciprocal process) from relationally emergent hybrid intelligence (a validated outcome that exceeds matched human-only and AI-only baselines). The six-month scientific case analysed here motivates the construct but does not establish emergence, becausematched counterfactual baselines were not collected. The framework operationalizes branch diversity, structural compression, cross-substrate coherence, validation debt, preventive rehearsal, post-hoc interface repair, and the temporal readability of epistemic drift. A three-clock model distinguishes the persistence of an interpretable trace, the migration rate of the shared state, and the observation time required for reliablevalidation. Formal expressions are treated as computational-level schemata; each is paired with observable proxies and independent-rating procedures. Coherence is explicitly separated from agreement, confidence, truth, and task performance. A spin-3/2 calculation illustrates how a global inconsistency signal can be converted into a discriminating recomputation protocol. A companion industrial-AI benchmark motivates a prevention–detection–repair–validation hierarchy, while its three-seed results are treated as directional rather than confirmatory.The model predicts an inverted-U relation between exploratory expansion and validated performance, benefits from preventive epistemic rehearsal, and productive disagreement when provenance and constraints remain aligned. It concludes with preregistrable studies using matched baselines, blinded adjudication, multiple models and seeds, independent scoring, andexternal verification. The proposal therefore treats hybrid intelligence not as fluent conversation,but as a governed relation whose claimed emergence must survive counterfactual and empirical tests.