AI Models Struggle With Sustained Focus in New Study

Researchers led by Suketu Patel published a study in PNAS Nexus identifying a significant flaw in modern artificial intelligence attention mechanisms ^[1]. While models like GPT-5 and Claude 4.1 excel at brief interactions, their accuracy drops sharply during long, complex tasks ^[1][2]. The team utilized the classic Stroop psychological test to measure this cognitive decline ^[2].

The Stroop task requires subjects to identify the color of a word when the text itself names a different color ^[3]. Although the AI models initially performed well, their ability to filter conflicting information worsened as the sequence continued ^[2][3]. This suggests that current transformer architectures lack the sustained focus found in human cognition ^[1].