When AI thinks like a brain

Artificial intelligence (AI) is advancing at breathtaking speed, but its foundations remain very different from those of the human brain. The most powerful artificial neural networks rely on billions of parameters and massive amounts of data to accomplish specific tasks. By contrast, the human brain learns from only a handful of examples, adapts rapidly to changing environments, and continues to function effectively in the face of uncertainty. Fast, efficient, and flexible, it embodies a form of intelligence that is profoundly contextual.

This contrast raises a fundamental question: can we envision an AI that not only mimics human performance but also draws directly from the brain’s own operating principles? To explore this, a team of American researchers from institutions including Rutgers, MIT, Yale, and the University of Nebraska–Lincoln took a decisive step forward. In a 2022 study published in Nature Communications, they demonstrated that it is possible to build computational models directly from human brain data, giving rise to a “neuro-inspired” AI that reflects the brain’s inner logic.

The neuro-inspired approach: beyond raw performance

Since their inception, artificial neural networks have been designed to maximize performance whether in image recognition, machine translation, or text prediction. Their efficiency is undeniable, but their rigidity poses a problem. Even slight changes in task conditions can destabilize them, whereas humans adjust seamlessly.

This vulnerability stems from how these systems are built. They do not adhere to the biological constraints of the brain limited connectivity, metabolic costs, or the particular topology of neural networks. To overcome these limitations, some researchers argue for a different path: integrating the brain’s own principles, not merely as a metaphor but as an actual architectural blueprint.

Here, the concept of functional connectivity becomes central. Functional MRI (fMRI) makes it possible to observe which brain regions are activated together and how they exchange information, both at rest and during complex tasks. This dynamic mapping much more than the anatomical layout of nerve fibers offers critical insight into the brain’s flexibility.

When the brain becomes a blueprint for AI

The American research team used this functional connectivity to construct what they call an Empirical Neural Network (ENN) an artificial neural network whose architecture is derived directly from human brain data. Instead of arbitrarily defining layers and connections, they identified, through fMRI, the regions of the brain responsible for different aspects of a cognitive task: sensory perception, contextual rules, and motor responses.

At the heart of their model were “conjunction hubs,” areas that integrate multiple streams of information: task rules, sensory stimuli, and expected responses. These hubs act as computational crossroads, allowing both the brain and the model to transform diverse inputs into adaptive behaviors.

The experiment centered on a cognitive control task called C-PRO (Concrete Permuted Rule Operations). Participants applied logical, sensory, and motor rules to various visual and auditory stimuli. Using fMRI data collected during this task, the researchers built a network capable of simulating activity flow between brain regions and predicting the expected motor outputs.

The results were striking. Without undergoing conventional training on millions of examples, the ENN successfully predicted participants’ responses at levels above chance. In other words, the architecture drawn directly from the human brain was enough to implement flexible computational transformations.

🔗 Explore further: Your Quantum Brain: Where Neuroscience Meets the Edge of Reality

Toward a more flexible, human-like AI

This kind of model represents a clear break from traditional AI. While classical networks are optimized for peak performance in narrow environments, ENNs prioritize adaptability and flexibility.

Two elements explain this difference:

Conjunction hubs proved essential. Removing them from the model caused performance to collapse, showing their central role in cognitive flexibility.
Nonlinearities were equally important. By incorporating functions similar to those observed in biological neurons, the researchers enabled the model to carry out conditional logic, critical for handling complex situations.

Although these models are not as fast as conventional AI systems, they excel in robustness. Faced with unexpected changes in context, they continue to function without breaking down. This ability to generalize so natural for the human brain points to a path toward AI that is not only smarter but also more resilient.

From AI to personalized medicine

The implications extend well beyond designing smarter AI. ENNs also open new possibilities for medicine and neuroscience.

By replicating the brain’s actual organization, ENNs could serve as virtual laboratories for testing clinical hypotheses. For instance, simulating the impact of a lesion within a functional network could help predict the outcomes of a stroke or neurodegenerative disorder.

In the future, this approach could even enable personalized medicine. By building a computational model from an individual’s fMRI data, clinicians could predict how their condition might progress, evaluate the potential effectiveness of a therapy, or tailor neuroprosthetics to their unique neural profile.

🔗 Discover more: The brain rot effect: Instant content and the erosion of our mental functions

The convergence of AI and neuroscience, then, is not merely a technological race it also offers new ways to understand the human brain and improve patient care.

Still, this close alignment between AI and the human brain raises profound questions. Should we aim to design machines that respect the biological logic of life, even if it blurs the line between natural and artificial?

An AI that “thinks like a brain” is not necessarily more powerful, but it is more contextual and human-like in its functioning. It values stability, learns from minimal examples, and resists disruption. These qualities are invaluable in a complex and uncertain world where raw performance is not always enough.

This perspective forces us to reconsider our goals: do we want AI to surpass humans in narrowly defined tasks, or should we create systems that resonate with our cognitive rhythms and modes of adaptation?