A new study finds that leading AI models struggle to maintain focus and suppress automatic responses as cognitive demands increase
Artificial intelligence models have become remarkably capable at understanding and generating language. But new research suggests that their impressive performance may conceal an important weakness: when faced with competing instructions and increasingly demanding tasks, they can struggle to maintain their focus and suppress an automatic response.
As explained here, a study published in PNAS Nexus examined whether transformer-based artificial intelligence systems possess something comparable to the executive control mechanisms that allow humans to stay focused, resolve conflicts, and override habitual responses.
The researchers found that two leading AI models—OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet—performed well on relatively simple attention tasks but showed substantial declines as the amount of conflicting information increased. The findings suggest that transformer-based systems may have important limitations when it comes to maintaining a goal and suppressing a strongly activated competing response.
Putting AI attention to the test
The research was led by Suketu Chandrakant Patel, a doctoral candidate in comparative and cognitive psychology at the City University of New York’s Graduate Center, working in Jin Fan’s laboratory at Queens College. Patel and his colleagues became interested in the question after observing the striking combination of strengths and weaknesses displayed by modern language models.
“When ChatGPT arrived, much of the excitement centered on its capacity for task completion, theory of mind, and emotional intelligence,” Patel said. “Yet it was also prone to hallucination and confabulation. LLM performance was strong on some tasks and surprisingly weak on others.”
The team wanted to investigate those limitations using a well-established psychological test rather than relying on conventional language benchmarks.
They chose the Stroop task, a classic experiment designed to measure how people deal with conflicting information.
What is the Stroop task?
The Stroop task dates back to the work of psychologist John Ridley Stroop in the 1930s. In the classic version, participants are shown a color word printed in either a matching or conflicting color.
For example, the word BLUE might appear in red ink.
If the instruction is to name the ink color, the correct response is “red.” But reading the word itself is highly practiced and automatic, creating competition between the two responses.
Humans typically become slower and make more mistakes when the word and its color conflict. This phenomenon is known as the Stroop effect.
The researchers were interested in whether AI models would show a similar pattern—and, more importantly, what would happen when the task became increasingly demanding.
“The Stroop task is fitting because the success of LLMs rests on the transformer’s attention mechanism,” Patel explained. “In humans, attention comprises three dissociable yet overlapping systems: alerting, orienting, and executive control. So we set out to test whether these models possess all three.”
Testing GPT-4o and Claude 3.5 Sonnet
The researchers tested OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet using visual versions of the Stroop task.
The models were shown images containing lists of words and were instructed either to read the words or identify the physical colors in which the words appeared.
The researchers varied the conditions. Some words were printed in matching colors, while others were deliberately mismatched. They also included neutral words and nonword controls consisting of strings such as “XXXX.”
Most importantly, the researchers varied the length of the lists—from just one word to as many as 40.
This allowed them to test not only whether the models could follow an instruction, but whether they could maintain that instruction as the task became longer and more demanding.
That ability is known in psychology as goal maintenance: keeping a task objective active while filtering out information that interferes with it.
Performance deteriorated as the lists grew
The models initially performed well.
With short lists, GPT-4o and Claude 3.5 Sonnet were generally able to identify the correct color even when the written word conflicted with it.
But performance deteriorated as the lists became longer.
GPT-4o’s accuracy on incongruent color-naming trials fell from 91% with five-word lists to 57% with 10 words, 22% with 20 words, and 15% with 40 words.
Claude 3.5 Sonnet proved somewhat more resilient. Its performance remained relatively stable through 20-word lists but then fell sharply, reaching 24% accuracy at 40 words in the incongruent condition.
The mixed trials produced a similar pattern. As the lists grew longer, both systems increasingly shifted away from the requested task of naming colors and toward their default behavior of reading the words.
The effect was particularly striking because the models could still read the words accurately.
In other words, the problem wasn’t simply that the systems became unable to process a long list. Instead, they appeared increasingly unable to suppress the competing response when reading and color naming came into conflict.
Why does this matter?
The researchers argue that the findings point to a distinction between two forms of attention.
Transformer models are extremely effective at selectively processing information. Their self-attention mechanisms assign different weights to different pieces of input, allowing them to identify relationships across long sequences.
But human attention involves more than simply selecting information.
Humans also have systems involved in executive control—the ability to maintain goals, resolve conflicts, adjust behavior, and suppress responses that are inappropriate for the current task.
According to the researchers, transformer attention lacks an obvious architectural equivalent of this kind of executive control.
“The central argument is that the limitation stems from the lack of an explicit mechanism for top-down modulation,” Patel explained.
In practical terms, the model can recognize the instruction and understand the task, but maintaining that instruction while suppressing a powerful competing response appears to become increasingly difficult as the task grows.
Understanding the difference between attention and control
This distinction is important because the word “attention” can mean different things when discussing humans and AI.
A transformer can direct computational resources toward certain tokens or pieces of information. This resembles one aspect of human attention known as orienting.
But executive control involves something more demanding.
Imagine telling a person:
“For the next several minutes, ignore everything you read and report only the color of the letters.”
The person has to continuously maintain that rule, notice when reading begins to interfere, and redirect their behavior.
The researchers argue that this kind of sustained, goal-directed control is not explicitly built into the standard transformer architecture.
Their results suggest that this difference becomes especially visible when the competing response is strong and the task continues for a longer period.
Could AI simply learn to overcome the problem?
Possibly.
The researchers are careful to acknowledge that their findings do not mean AI models are permanently incapable of performing the Stroop task.
With enough training examples, a model could potentially learn the specific patterns required to perform it successfully.
But the researchers argue that this would not necessarily demonstrate the emergence of a general-purpose executive control mechanism.
A model might learn the solution through extensive exposure to the particular task rather than developing a flexible ability to suppress competing responses across unfamiliar situations.
That distinction matters when thinking about artificial general intelligence.
What about reasoning and code?
Modern AI systems can sometimes use external tools, code generation, or additional reasoning processes to solve difficult tasks.
The researchers argue that such strategies can complicate cognitive tests like the Stroop task.
For example, if a model uses code to extract the colors from an image instead of actually resolving the conflict between the written word and its color, it may produce the correct answer—but it has effectively bypassed the cognitive challenge the experiment was designed to measure.
Humans can do something similar. A person could deliberately blur their vision, cover the word, or use another strategy to avoid reading it. Such an approach might produce the correct answer, but it would no longer be a clean test of inhibitory control.
The findings don’t prove that AI cannot develop executive control
The researchers’ conclusions should therefore be interpreted carefully.
The experiment demonstrates a significant and reproducible weakness in the tested models under the specific conditions of the Stroop task. It does not, by itself, prove that transformer-based AI can never develop executive control or that artificial general intelligence is impossible without a brain-like architecture.
The researchers themselves acknowledge that further training could improve performance.
Their broader claim is more structural: today’s transformer architecture may lack mechanisms that correspond closely to the way human executive control maintains priorities, resolves conflicts, and adapts behavior over time.
That raises an important question for AI development:
Will these capabilities emerge naturally as models become larger and are trained on more data, or will future systems require new architectural mechanisms specifically designed for sustained cognitive control?
A challenge for the path toward AGI
The question is particularly important because the AI industry has increasingly focused on scaling models, expanding context windows, and giving systems access to external tools.
Those approaches have produced enormous gains in capability.
But the researchers argue that simply processing more information is not necessarily the same as developing the ability to control how that information is processed.
For an AI system expected to operate autonomously over long periods, that distinction could become crucial.
An advanced system may need to maintain a complex objective while dealing with distractions, conflicting instructions, changing circumstances, and its own previously generated mistakes.
That requires more than recognizing information. It requires maintaining priorities and continuously deciding which competing signals should win.
What comes next?
The researchers say future AI architectures may need more explicit mechanisms for executive control, including systems capable of maintaining goals across extended interactions, resolving conflicts between competing computations, and dynamically adjusting attention.
The study therefore raises a broader question about the future of artificial intelligence.
Today’s models demonstrate extraordinary abilities in language, perception, reasoning, and tool use. But their ability to sustain a goal under pressure may not yet resemble the flexible executive control seen in humans.
Whether that limitation can be solved through continued scaling and training—or whether it will require fundamentally different AI architectures—remains an open question.
The full study, Deficient executive control in transformer attention, was authored by Suketu Chandrakant Patel, Hongbin Wang, and Jin Fan and published in PNAS Nexus in June 2026.
