WP_AI_8.7 AI Capability Proportionality

This paper introduces AI Capability Proportionality, a two-sided governance framework for determining how much AI capability should be developed, exposed, deployed, purchased, and maintained in relation to real need. The paper argues that capability availability does not create capability necessity. On the supply and deployment side, AI systems should expose only the Effective Power justified by the task, authority, environment, and possible consequences. On the demand and user side, enterprises and institutions should acquire and upgrade AI according to actual operational need and reasonably foreseeable development, rather than continuously pursuing greater capability simply because more powerful systems become available. The framework distinguishes cognitive capability from operational capability, defining cognitive capability as a multiplier of the tools, permissions, connectivity, persistence, autonomy, and environmental reach available to a deployed system. It develops the concepts of Minimum Sufficient Capability, Justified Need, Capability Reserve Margin, Excess Capability Exposure, Capability Overinvestment, and Capability Creep. The paper also examines the economic and social consequences of disproportionate AI deployment. At the enterprise level, unnecessary high-capability deployment can increase integration, migration, security, staffing, energy, infrastructure, and governance costs while crowding out other productive investment. At the social level, repeated capability pursuit can influence expectations, capital allocation, infrastructure expansion, and broader resource concentration, creating conditions in which AI capability continues to grow while its marginal or net social benefit declines. A central distinction is made between falling consumer-level access costs and the total organizational cost of enterprise- and industrial-scale AI deployment. The paper further proposes a three-zone model of social benefit and a two-sided proportionality test for evaluating whether Effective Power remains aligned with Justified Need. Within the broader WP_AI_8: Foundations of AI Safety and Governance series, this paper extends earlier work on capability release, governance objects, and black-box governance by asking a further question: Even when AI capability can be developed, released, and governed, how much of that capability is actually necessary? The paper concludes that AI progress should not be measured by capability growth alone, but by the degree to which capability remains proportionate to real task demand, enterprise capacity, governance requirements, and social net benefit.