level: research
classical cognition views problem solving as exploratory search through structured spaces, where repeated attempts gradually compress efficient representations. predictive ai systems change this by offering solutions before internal exploration begins, creating a regime where stabilization precedes diversification. this paper introduces a geometric dynamical framework to study how attention moves through strategy landscapes under three forces: stabilizing drift, endogenous exploratory perturbation, and responsiveness-gated learning.
the model treats predictive assistance as exogenous exploratory compression. it stabilizes decision trajectories early, before self-generated exploration can broaden the accessible strategy space. the framework produces three key findings. sustained predictive stabilization reduces the diversity of explored strategies over time. it also narrows the basin of attraction around assisted trajectories, making users less likely to discover alternative approaches. finally, the timing of assistance matters: early intervention locks in narrow paths, while later assistance allows more exploration.
these results suggest that predictive ai can inadvertently limit human problem-solving flexibility. by compressing search too soon, it may prevent the discovery of novel or better strategies. the geometric approach provides a formal way to analyze such effects, offering insights for designing ai systems that balance helpful stabilization with the need for exploratory learning. future work could extend the model to multi-agent settings or test predictions in human-ai interaction experiments.
why it matters: understanding how predictive ai shapes exploration helps designers avoid systems that prematurely narrow user strategies, preserving human creativity and adaptability.