Large language models exhibit a critical vulnerability in executive control, failing the classic Stroop task with severity that increases alongside input length. A study released on January 10, 2025, and published in PNAS Nexus by researchers Suketu Chandrakant Patel, Hongbin Wang, and Jin Fan demonstrates a fundamental dissociation between task recognition and task execution in transformer-based architectures. The Stroop task serves as a clinical assessment of human executive function, requiring participants to identify the ink color of printed words while suppressing the automatic impulse to read the text. While humans maintain high accuracy and stable performance across extended word lists, recent testing reveals that major commercial AI models struggle to inhibit word-reading biases. The research team evaluated GPT-4o, Claude 3.5 Sonnet, GPT-5, Claude Opus 4.1, and Gemini 2.5 on incongruent word-color pairs provided without explicit task instructions. Performance metrics showed a steep accuracy decline as list length increased. GPT-4o began at ninety-one percent accuracy on a five-word list but fell to fifty-seven percent at ten words and plummeted to fifteen percent at forty words. Claude 3.5 Sonnet maintained stability through twenty words before collapsing to twenty-four percent accuracy at the forty-word threshold. When tested on mixed lists containing both congruent and incongruent stimuli, all evaluated models suffered near-total performance failure on mismatched items, frequently defaulting to reading the word rather than naming the ink color. The authors attribute this degradation to architectural limitations inherent in current attention mechanisms. While transformer models demonstrate the capacity to identify the structural requirements of the Stroop paradigm, they lack the dynamic conflict-resolution pathways necessary to override learned linguistic priors. Human cognition successfully suppresses automatic word recognition through top-down executive control, allowing sustained focus on color identification. In contrast, AI systems process sequential tokens without an equivalent inhibition layer, causing attention to drift as context windows expand. This performance collapse underscores a persistent gap between symbolic task comprehension and functional execution in artificial intelligence research. The findings suggest that current language models excel at pattern matching and instruction following within constrained contexts but remain vulnerable to cognitive load when real-time decision-making requires active response suppression. Future architectural developments may require dedicated inhibition modules or hybrid reasoning frameworks to replicate biological attention dynamics. The research conducted by Patel, Wang, and Fan provides empirical evidence that transformer-based systems lack the autonomous regulatory mechanisms required for complex executive tasks. As AI systems are deployed in environments demanding sustained focus and adaptive filtering, understanding these attentional boundaries will be critical for risk mitigation and model refinement.