Proximal State Nudging: Reducing Skill Atrophy from AI Assistance

Megha Srivastava, Jonathan Ouyang, Eric Ziyang Zhou, Andrew Silva, Emily Sumner, Dorsa Sadigh, Yuchen Cui, Deepak Edakkattil Gopinath, Guy Rosman

Keywords: cooperative ai, human-ai interaction, shared autonomy, education, skill retention, learning-aware assistance, human-centered AI, AI safety, trustworthy AI, shared control, reinforcement learning

TL;DR: We introduce a cooperative human-AI shared autonomy framework, Proximal State Nudging, that maintains safety and task performance while actively supporting long-term human skill development through learning-aware assistance.

Abstract: The gradual loss of human skills, or skill atrophy, is a rising concern as users increasingly rely on AI assistance. This problem is particularly salient in cooperative AI systems, such as aircraft piloting or driving, where humans and AI agents jointly share control and decision-making. In these settings, human operators often struggle to disentangle which outcomes arise from AI intervention versus their own actions, undermining opportunities for their own learning and long-term skill retention. In this work, we propose \textsc{Proximal State Nudging}, a cooperative shared-control algorithm that balances assistance with human skill development. Rather than optimizing solely for combined Human+AI task reward, our method also gradually ``nudges'' human users toward states in the environment where they are most likely to improve their own, unassisted competence. In two simulated environments (Discrete MDP and LunarLander), Proximal State Nudging outperforms existing shared autonomy baselines in improving a student's unassisted performance. We further validate our approach through two human subject studies (Parallel Parking and High Performance Racing, n=60) using the high-fidelity CARLA driving simulator, showing that we can build real-world cooperative AI systems that support human agency and skill retention without sacrificing performance.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 186