Even GPT-5 Failed This Human Attention Test

A recent scientific study has identified a significant limitation in the cognitive capabilities of advanced artificial intelligence. According to reports from SciTechDaily and Caliber.Az, these AI models, including the high-profile GPT-5, have failed a specific human attention test known as the Stroop test. The findings indicate that the performance of these models degrades specifically as the length of the task increases, suggesting a failure to maintain a consistent level of attention or processing accuracy over extended sequences of information. Coverage from Baku.ws emphasizes that scientists have now identified what they describe as the main weakness of artificial intelligence through this specific testing methodology.

SciTechDaily further highlights that the failure occurs even in the most advanced iterations of large language models, such as GPT-5, which were previously expected to handle such cognitive tasks with greater ease. Caliber.Az provides the specific detail that the failure is directly linked to the increasing length of the task, meaning the models struggle more as the sequence of the Stroop test grows longer. To understand why this matters, it is necessary to recognize that the Stroop test is a standard psychological measure used to evaluate human attention and processing speed. It typically requires a subject to ignore a distracting stimulus to focus on a specific task, such as naming the color of a word when the word itself spells a different color.

The fact that AI models fail this test as the task expands indicates a fundamental gap between synthetic processing and human-like attention mechanisms, which has broader implications for how AI handles complex, long-form cognitive instructions. Moving forward, observers will be looking for how researchers address this identified weakness in AI architecture. Based on the reports from Caliber.Az, Baku.ws, and SciTechDaily, the primary focus remains on the correlation between task length and model failure. Future developments will likely center on whether subsequent updates to GPT-5 or other models can overcome this specific attention barrier or if this limitation remains an inherent characteristic of current artificial intelligence design.