Title: Effective AI Governance Must Account for Capabilities Beyond the Model Itself

Abstract

Current approaches to governing frontier artificial intelligence predominantly rely on a model-centric framework, predicated on the assumption that a system’s capabilities are determined chiefly by the computational power and datasets utilized during the training phase. This position paper contends that such model-level governance loses its efficacy as technological progress becomes increasingly fueled by "non-model gains"—advancements that operate independently of the core model’s development. We define this concept and introduce a taxonomy categorizing capability enhancements into three specific vectors: inference gain, which involves scaling computational resources during testing; systems gain, encompassing post-training improvements like scaffolding; and asset gain, which refers to augmenting models with restricted resources. The paper illustrates how these vectors, in conjunction with potential future developments in embodiment, continuous learning, and AI diffusion, threaten to compromise risk management protocols that depend heavily on pre-deployment assessments and mitigations. Furthermore, we survey governance strategies that transcend the model level, including system, entity, agent, and cloud governance. Finally, we underscore the critical role of societal resilience as a necessary complement to these governance layers.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC