In a new study published in PNAS Nexus, scientists have demonstrated that artificial intelligence can predict different types of human intelligence by analyzing connections in the brain. Using neuroimaging data from hundreds of healthy adults, they found that predictions were most accurate for general intelligence, followed by crystallized intelligence, and then fluid intelligence. The findings shed light on the distributed and dynamic nature of intelligence, demonstrating that it arises from the global interplay of brain networks rather than isolated regions.
While prior research has established that intelligence is not localized to a single brain region but rather involves distributed networks, many studies have relied on traditional methods that focus on isolated brain features. These approaches have offered limited insights into how intelligence arises from the interplay of brain structure and function. By employing machine learning to analyze brain connectivity, the researchers aimed to overcome these limitations.
A key focus of the study was the distinction between three major forms of intelligence: general, fluid, and crystallized. General intelligence, often referred to as “g,” is a broad measure of cognitive ability that encompasses reasoning, problem-solving, and learning across a variety of contexts. It serves as an overarching factor, capturing shared elements between specific cognitive skills.
Fluid intelligence, a subset of general intelligence, refers to the capacity to reason and solve novel problems without relying on prior knowledge or experience. This type of intelligence is often associated with abstract thinking, pattern recognition, and adaptability. In contrast, crystallized intelligence represents the ability to use knowledge and skills acquired through education, culture, and experience. It includes abilities such as vocabulary, reading comprehension, and factual knowledge.
“Our research group aimed to investigate how individual differences in intelligence, or general cognitive ability, are manifested in the human brain. We are convinced that the interconnections between different brain regions—believed to reflect communication pathways—play a particularly critical role,” said study author Kirsten Hilger, the head of the Networks of Behavior and Cognition research group at Julius-Maximilians-Universität Würzburg.
“Many studies published in recent years have predicted individual differences in intelligence based on these communication pathways, known as functional brain connectivity. However, the primary goal of these studies has often been to achieve the highest possible prediction performance, while insights into the concept of intelligence and the question of how intelligence may arise from these communication pathways were largely absent.”
“With our study, we aim to address this limitation by providing methods and approaches to gain interpretable insights into the concept of intelligence, i.e., to actually learn something about how intelligence evolves from the brain,” she explained.
To predict intelligence, the researchers used data from the Human Connectome Project. Their analysis included 806 participants aged 22 to 37, who were free of cognitive impairments. Brain connectivity was assessed using functional magnetic resonance imaging (fMRI) during resting states and seven tasks designed to activate different cognitive processes, such as working memory, language, and emotional recognition. Fluid intelligence was measured using tests independent of prior knowledge. Crystallized intelligence was assessed with vocabulary and reading tasks. General intelligence, combining elements of both, was calculated as a composite score.
The researchers trained machine learning models to analyze connections between 100 defined brain regions across eight cognitive states. They compared models using connections proposed by leading intelligence theories with those trained on randomly chosen connections. Additionally, they applied a technique called relevance propagation to identify which brain connections most contributed to the predictions.
Among the different types of intelligence examined, general intelligence was the most accurately predicted by the machine learning models. This finding suggests that general intelligence, as an overarching cognitive ability, may be associated with more consistent or prominent patterns of brain connectivity compared to the other types. Crystallized intelligence was also predicted with considerable accuracy, while predictions for fluid intelligence were less precise.
One of the key insights was that brain activity during cognitively demanding tasks yielded more accurate predictions of intelligence than activity during resting states. Tasks that required working memory or language processing, for instance, significantly improved the models’ ability to predict fluid and general intelligence. This finding highlights the dynamic nature of brain connectivity and its importance in supporting higher-level cognitive processes.
In contrast, crystallized intelligence, which is tied to long-term knowledge and skills, appeared to rely more on stable, task-independent brain networks. The use of latent connectivity measures, which integrate information across multiple brain states, further enhanced predictions of crystallized intelligence, suggesting that this form of intelligence may emerge from widespread, stable communication patterns in the brain.
The researchers also found that models that incorporated connections between brain regions highlighted by theories like the parieto-frontal integration theory outperformed those trained on randomly selected regions. This reinforces the idea that certain brain networks, particularly those involving the prefrontal and parietal areas, are integral to cognitive functioning. However, whole-brain models consistently outperformed theory-driven models, indicating that intelligence likely arises from a more extensive and distributed network of connections than previously understood.
“Individual differences in intelligence are not manifested in few circumscribed regions in the brain, but instead in a communication mechanism involving the whole brain,” Hilger told PsyPost. “Previous neurocognitive models of intelligence are not wrong, but need to be extended towards the inclusion of the whole brain and more focus on the mechanisms instead of specific brain regions.”
The analysis identified approximately 1,000 specific brain connections as being most predictive of intelligence. These connections were not confined to isolated regions but were distributed across the brain, involving major networks such as the default mode network, the frontoparietal control network, and the attention networks. These findings underscore the idea that intelligence is a global property of the brain rather than the product of activity within a single region or system.
Interestingly, the researchers found that the brain’s ability to compensate for missing connections was remarkably high. Even when entire networks were excluded from the models, predictions of intelligence were only minimally affected.
“Artificial removal of complete large-scale functional brain systems affects predictive performance surprisingly little,” Hilger said. “Thus, there seem to be some redundancy in the neural code of intelligence differences.”
While the study provided significant insights into the neural underpinnings of intelligence, it is not without limitations. One notable constraint is the narrow age range of participants, which included only healthy adults aged 22 to 37. This limits the generalizability of the findings across the lifespan, particularly for children and older adults.
In addition, although the study identified approximately 1,000 brain connections as most predictive of intelligence, the exact nature of these connections and their functional roles remain unclear. Investigating the specific processes these connections support—such as memory, attention, or executive control—could help clarify how they contribute to different types of intelligence. Furthermore, exploring individual differences in neural strategies for problem-solving and knowledge application might illuminate why some connections are more predictive than others.
“In sum, our results suggest intelligence as emerging from global brain characteristics, rather than from isolated brain regions or single neural networks,” the researchers concluded. “In a broader context, our study offers a framework for future predictive modeling studies that prioritize meaningful insights into human complex traits over the mere maximization of prediction performance.”
The study, “Choosing explanation over performance: Insights from machine learning-based prediction of human intelligence from brain connectivity,” was authored by Jonas A. Thiele, Joshua Faskowitz, Olaf Sporns, and Kirsten Hilger.
