Governing Human–AI Systems: A Co-Agency Equilibrium Model for Dynamic Authority Allocation and Accountability

Human–AI decision authority is analyzed in relation to changes in relative competence and decision risk, taking institutional oversight capacity into account. The proposed Co-Agency Equilibrium Theory (CAET) and Co-Agency Equilibrium Model (CAEM) offer a sociotechnical perspective on humans and AI working as co-agents with varying shares of decision authority. CAET is a normative design framework connecting relative competence, moral salience, irreversibility and independent oversight capacity with human authority and a normalized risk–oversight gap. CAEM combines a logistic policy surface, accountability gap feedback and a gradual adjustment of the human control level, with a switch to full human control if moral salience or irreversibility reaches or exceeds the corresponding threshold. The decision to switch is implemented in the following episode. The fixed point is also a unique point in cases of constant external inputs and convergence is achieved when the computed sufficient condition is satisfied. The authority boundary identifies whether the risk that is ceded to AI is within oversight capacity. Simulations use matched inputs for the feedback and no-feedback cases, covering coupled risks, changing capacity, noise and alternative risk functions. Even after the allocation has stabilized, a positive gap may remain due to a lack of sufficient capacity to oversee. In the example with time-varying capacity, feedback cut the cumulative positive gap by 6.4%; this was accompanied by more time with a positive gap and greater variation in authority. Delay in takeover means there is an extended period of exposure before people have effective control. These conclusions concern the model and its assumptions. The findings are not empirical validation, nor able to guarantee accountability or even safety.