Stroop Test Reveals AI’s Hidden Attention Gap
New research in PNAS Nexus finds that advanced language models like GPT-4o and Claude 3.5 Sonnet can handle language but falter on a Stroop-style test as context length increases: their accuracy in naming ink color collapses from near-perfect to single digits (GPT-4o from 91% to 1%; Claude to about 10%), because current transformer architectures lack explicit top-down executive control to override automatic word-reading. The study argues scaffolding or tool-use cannot substitute true cognitive control, and that achieving artificial general intelligence will likely require integrating executive-control mechanisms directly into AI architectures rather than relying on scale alone.