Abstract
This paper introduces a model of governed distributed cognition in long-horizon human–AI collaboration systems. It addresses a central gap in existing research: while prior work has examined either human cognition, artificial intelligence capabilities, or loosely coupled distributed cognition, there is limited understanding of how stable reasoning emerges across sustained human–AI interaction. The paper advances the core claim that cognition in such systems is not located within the human or the AI model alone, but instead emerges as a distributed, system-level process across three components: human judgment, AI-generated reasoning, and persistent artifact-based memory. Crucially, this cognitive process is not free-form. It is regulated through governance mechanisms that structure how reasoning unfolds over time. Building on prior work that formalizes governance architecture for human–AI collaboration, this paper extends the framework by introducing three defining properties of cognition in governed systems: Distributed cognition — reasoning is partitioned across human, AI, and artifacts, with no single component containing the full process Governed cognition — the interaction is structured by constraints, protocols, and linguistic control signals that regulate reasoning Recoverable cognition — the system incorporates intrinsic mechanisms for detecting, diagnosing, and repairing errors during the reasoning process The paper further formalizes a governed cognitive loop, in which cognition unfolds as a recurring cycle of intent formation, generative expansion, state anchoring, evaluation, drift detection, and repair. Within this loop, error is not treated as a breakdown but as an integral component of a self-correcting cognitive system. Empirically grounded in a longitudinal human–AI collaboration, the framework demonstrates that stability in long-horizon reasoning is not achieved by eliminating error, but by embedding structured mechanisms for correction and continuity within the cognitive process itself. The paper contributes a new theoretical model of cognition for human–AI systems, reframing cognition as a distributed, regulated, and recoverable system-level phenomenon. This perspective has implications for AI research, human–AI collaboration design, and the development of reliable long-horizon reasoning systems.