DEMO
Sandeep Banik and Naira Hovakimyan
As autonomous agents integrate into everyday processes—assistive robotics, autonomous driving, aviation, and co-working AI—a critical question emerges: how should authority be dynamically allocated between humans and AI?
Existing approaches remain fragmented. Robotic systems rely on blending heuristics; AI assistants employ implicit delegation; agentic systems lack principled authority structures. These methods fail to account for the co-evolving dynamics of human-AI collaboration.
This work introduces Flip-Team, a game-theoretic framework for authority switching that extends to learning-based settings where dynamics or costs are unknown. Under cooperative assumptions, authority switching becomes an identical-interest dynamic game with closed-form solutions for linear-quadratic systems. For complex nonlinear dynamics, we learn value functions from interaction data: authority transfers when expected cost under AI control exceeds that under human control by more than the switching penalty. Parameterizing value functions as neural networks yields policies that generalize without explicit system identification.
We extend beyond identical interests by formulating switching as a potential game, capturing scenarios where human and AI objectives partially diverge while admitting tractable equilibrium analysis. Applications span AI-assisted decision support, conversational turn-taking, and collaborative content generation. Preliminary results demonstrate learned switching policies outperform fixed-authority baselines, providing a unified framework where authority adapts to context, competence, and cost.