Generative AI as a Third Agent: Rethinking Academic Work and Principal-Agent Dynamics in Higher Education

Abstract

The rapid diffusion of generative AI in higher education is transforming academic work in ways that extend beyond technical efficiency or pedagogical innovation, reaching the foundations of institutional authority, human agency, and educational purpose. This paper develops a nested principal–agent model in which generative AI operates as a third, non-human actor that simultaneously alters the production and observability of academic work across two layers: university–professor and professor–student. Within this framework, we introduce the concept of cascading moral hazard: the structural condition whereby delegation in the upper layer weakens the principal’s capacity to monitor action in the layer below. Conditional on two maintained assumptions—a synchronic monitoring condition and a diachronic cognitive-dulling hypothesis, according to which repeated delegation dulls the very capacity being delegated, leaving it dormant rather than lost—the model yields a Nash equilibrium in which delegation rises across both layers, the informational value of credentials declines, and no actor has an individual incentive to deviate. These properties are formal implications of the model; the dulling hypothesis is an assumption grounded in embodied cognition (Varela, Thompson, and Rosch, 1991) and the cognitive-offloading literature, not an empirical finding of the present study. An exploratory pilot survey of fifty-two faculty members and students at a public university in Cusco, Peru, fielded between January and April 2026, provides cross-sectional evidence consistent with the model’s synchronic implications: higher self-reported delegation is associated with a larger perceived signal–competence gap. We argue that the institutional response required is not stricter rules or better detection technology, but a redesign of the principal–agent structures themselves.