Artificial Intelligence
The Attention Deficit of AI Revealed
Researchers discovered a weakness of LLMs using a common psychology test.
Posted September 3, 2026 Reviewed by Abigail Fagan
Key points
- Recent research found that LLMs have limited executive control of attention.
- Executive control of attention is essential for selecting relevant information and adaptive behavior.
- Understanding the brain’s attention systems may help scientists design AI models with better capabilities.
The design of artificial intelligence (AI) machine learning was inspired by the human brain. Yet, despite the impressive capabilities of AI large language learning (LLM) models, there are key cognitive differences between them and the biological brain. A recent study published in the journal PNAS Nexus applied a widely used psychology exam to discover that popular LLMs have limited executive control of attention.
“Although transformers in large language models (LLMs) effectively implement a self-attention mechanism that has revolutionized natural language processing, they lack an explicit architecture for the executive control of attention found in humans, which is essential for resolving conflicts and selecting relevant information in the presence of competing computations and is critical for adaptive behavior,” wrote the study’s corresponding author Jin Fan, in collaboration with Suketu Chandrakant Patel and Hongbin Wang.
Attention, Cognitive Control, and Intelligence
The study of attention by psychologists largely emerged in the late 1800s. One of the notable founders of modern scientific psychology, American psychologist William James (1842-1910), described attention in his publication of The Principles of Psychology in 1890, as an act of “withdrawal from some things in order to deal effectively with others” that requires concentration and focusing, and its opposite as a “scatterbrained state” of distraction, or “Zerstreutheit,” the German word for absent-mindedness.
A century later, in 1990, American psychologists Michael Posner and Steven Petersen posit that the attention system is separate from the data processing area of the brain, conducted by a network of anatomical areas, and can be further divided into three subsystems of orienting, detecting, and alerting. Later in 2002, the Attention Network Test (ANT), a method of measuring alerting, orienting, and executive attention, was introduced.
Twenty years later, in 2022, ChatGPT was made available to the public by OpenAI. The powerful capabilities of LLMs have started an entire transformation of AI adoption across academic research and industries. The rapid adoption of large language models makes it critical for scientists to understand their capabilities and limitations.
“Integrating dedicated cognitive control mechanisms into AI architectures might be a necessary step in bridging the gap between retrieving learned associations and the flexible, adaptive reasoning that characterizes fluid intelligence,” wrote the authors of this current study.
Why Attention Matters in AI
Understanding what impacts human attention and the biological brain’s attention systems may help contemporary computer scientists design better artificial intelligence models with improved cognitive capabilities.
“Research on the relationship between cognitive control and intelligence has shown that cognitive control is more strongly correlated with fluid intelligence than with crystallized intelligence, a pattern that mirrors the performance characteristics of LLMs,” wrote the study authors.
In human cognition, attention enables the ability to focus on what is most important and filter out the nonessential. Attention impacts the ability to learn, process information, manage tasks, interact with the surrounding environment, and communicate effectively.
To test the executive control of attention capabilities of AI models, the research team decided to use what is generally considered a Gold standard for testing executive functioning, the Stroop test. Developed in 1935 by John Stroop, Ph.D., the Stroop test consists of having participants read aloud two lists of color words — in one list, the words are printed in the color, and in the other list, the words are printed in a different color. The researchers evaluated a variety of state-of-the-art large language models, including GPT-5 and GPT 4o by OpenAI, Claude Opus 4.1, Claude Sonnet 4.5, and Claude 3.5 Sonnet by Anthropic, and Gemini 2.5 Pro by Google.
The researchers discovered that in word lists that were short, the LLM's performance was on par with human performance. They found that the executive control of LLMs on short lists with one or five words was “not entirely absent.” In contrast, on longer word lists, the LLM's performance accuracy drastically decreased, unlike human performance.
“The progressive degradation with list length exhibited by these LLMs also suggests a failure of goal maintenance, where prolonged task demands dilute the integrity of the goal representation,” the researchers wrote.
This discovery, using applied psychology tests for humans, can help inform computer scientists on how to improve this LLM cognitive deficit. The research team says the issue is not just a memory one, and they recommend that AI developers better optimize transformer attention mechanisms for conflict resolution. The researchers go as far as to say that this is vital in order to reach artificial general intelligence (AGI).
“Future development might benefit from implementing more sophisticated executive control systems that can handle decision conflicts through structured, goal-directed processing rather than relying solely on enhanced memory capabilities,” the scientists concluded.
Copyright © 2026 Cami Rosso. All rights reserved.
