New study: AI should follow high-performing humans, not lead

A new research paper from arXiv (2504.20903) challenges the prevailing wisdom that AI should always lead in decision-making. The authors, Prothit Sen and Sai Mihir Jakkaraju, develop a computational model of joint sequential adaptation where two agents differ in a single key characteristic: memory regime. Human agents operate under a recency-weighted regime that privileges recent outcomes, while AI agents use a uniform-memory regime that weights a window of past outcomes equally. The model varies task complexity (N), within-task coupling (K), and cross-agent coupling (C) across modular and sequential task structures.

The results reveal three key mechanisms: threshold dynamics create absorbing high- and low-payoff regimes; uniform memory amplifies inherited trajectories (good or bad) while recency weighting corrects locally but is volatile at scale; and memoryless stochastic search acts as an escape from poor trajectories. The study concludes that joint performance is maximized not by "AI-first" design, but when scale-free AI adaptation follows a high-performing human. The paper includes experimental validation and offers organization designers a contingency logic for human-AI collaboration.