A recent study published in PNAS Nexus on June 14, 2026, reveals that advanced large language models (LLMs) like GPT-5, Claude 3.5 Sonnet, and Gemini 2.5 struggle with sustained attention when subjected to the classic Stroop task. While these models perform well on short, simple sequences, their accuracy collapses significantly as task complexity and list length increase, indicating a fundamental difference between AI processing and human executive control. The researchers, led by Suketu Patel, argue that AI systems lack the biological capacity to suppress automatic responses—such as reading a word instead of identifying its ink color—over extended periods. This dissociation between task recognition and execution suggests that current transformer-based architectures do not replicate the human brain's ability to maintain goal-oriented focus under cognitive conflict. The study highlights a critical limitation in modern AI: the inability to maintain long-term instruction adherence when faced with competing information. As task length grows, these models revert to their training biases, prioritizing word recognition over the specific task instructions, leading to near-zero accuracy in complex, mismatched scenarios. For users and developers, this research serves as a warning that impressive performance in reasoning or coding does not equate to human-like cognitive stability. Future AI applications requiring sustained focus and inhibition of automatic responses may need architectural improvements to overcome these inherent attention deficits.
