Study Queries Centaur AI’s Human-Level Cognition

A new study published in National Science Open challenges recent claims regarding the Centaur AI model’s cognitive abilities ^[1]. Although the model reportedly achieved human-level performance across 160 different tasks, researchers suggest these results might not stem from genuine understanding ^[1]. Instead, the model may simply be overfitting and memorizing specific data patterns ^[1].

The research team argues that achieving high scores on benchmarks does not always indicate human-like thinking ^[1]. By analyzing the model's responses, they identified signs that suggest pattern recognition rather than complex reasoning ^[1]. This distinction remains vital for future AI development and accurately measuring progress toward general intelligence ^[1].