Alignment between Brains and AI: Evidence for Convergent Evolution across Modalities, Scales and Training Trajectories

摘要

Artificial and biological systems may evolve similar computational solutions despite fundamental differences in architecture and learning mechanisms—a form of convergent evolution. We provide large-scale evidence for this phenomenon through comprehensive analysis of alignment between human brain activity and internal representations across over 600 AI models spanning language and vision domains (1.33M to 72B parameters). Analysis of 60 million alignment measurements reveals that higher-performing models spontaneously develop stronger brain correspondence without explicit neural constraints, with language models demonstrating markedly stronger correlations ($r=0.89, p<7.5 \times 10^{-13}$) than vision models ($r=0.53, p<2.0 \times 10^{-44}$). Crucially, longitudinal training analysis shows that brain alignment consistently emerges prior to performance improvements, suggesting that developing brain-like representations constitutes a fundamental stepping stone toward enhanced capabilities. We identify systematic organizational patterns reflecting human cognitive architecture: language models exhibit strongest alignment with limbic and integrative regions, while vision models show progressive correspondence with visual cortical hierarchies. These findings establish that optimization for task performance naturally drives AI systems toward human-like computational strategies.