Steven Weisberg, a researcher at The University of Texas at Arlington, put some of the most advanced artificial intelligence tools available to work on structural brain scans and found they could not predict how well a person finds their way around. The study, which tested multiple deep learning architectures on MRI data from approximately 90 healthy young adults, returned only weak results, challenging a decades-old assumption that the size or shape of brain regions like the hippocampus can explain why some people are better navigators than others. The findings, detailed in a university news release, land at a moment when AI is being rapidly adopted across neuroscience, raising hard questions about what structural imaging can actually tell us about everyday human abilities.
Multiple AI Models, Minimal Predictive Power
The research team trained three distinct classes of machine learning models on T1-weighted structural MRI data: 3D convolutional neural networks, graph-based convolutional neural networks, and radiomics feature sets paired with ensemble classifiers. Each approach was designed to extract different kinds of anatomical information from brain scans, from raw voxel patterns to graph-theoretic summaries of cortical geometry. The dataset came from an earlier preregistered study of roughly 90 participants, all healthy young adults who had completed spatial navigation tasks, giving the AI models a well-characterized behavioral benchmark to predict.
Despite the variety of methods, every model showed only weak ability to forecast individual differences in navigation behavior from brain structure alone. That outcome is striking because deep learning tools are often pitched as capable of detecting subtle patterns that traditional statistical tests miss. If hidden structural signatures of navigation skill existed in these scans, at least one of the three model families should have picked up a reliable signal. The fact that none did suggests the problem is not a lack of analytical power but a genuine absence of strong structural markers in the typical adult brain, a point emphasized in commentary from Weisberg’s team that frames the result not as a failure of AI but as evidence that the field needs to look beyond gross anatomy for answers.
Training Changes Behavior, Not Brain Structure
A separate line of evidence strengthens the case that structural volume is the wrong place to look. In a randomized, multiweek training intervention, participants practiced navigation and verbal memory tasks while researchers tracked them with both structural and functional MRI before and after training. Over time, the trainees became significantly better at the skills they practiced, showing faster learning curves and clear behavioral improvement.
Yet when investigators examined the post-training structural scans, they found no detectable changes in overall brain anatomy, including no shifts in hippocampal volume or hippocampal subfield measurements. Instead, what changed was brain function: the same project, reported in a longitudinal imaging study, showed that functional activation patterns and network-level connectivity reorganized in response to training even as the physical structure of the brain remained stable. This dissociation between behavior and anatomy is directly relevant to the AI study’s null result. If weeks of intensive practice can make someone a better navigator without altering the tissue that structural MRI captures, then no algorithm trained on a single snapshot of brain anatomy should be expected to decode navigation ability with high accuracy.
The Taxi Driver Legacy and Its Limits
The expectation that brain structure should predict navigation skill traces back to a landmark study of licensed London taxi drivers. That work, published in a high-profile journal, found that cabbies who spent years memorizing the city’s labyrinthine street network had a larger posterior hippocampus compared to matched controls, with anterior hippocampal differences running in the opposite direction. The volume differences correlated with time spent driving a taxi, creating a compelling narrative: intense spatial practice physically reshapes the brain, and those structural changes track expertise.
But that narrative has not held up well outside expert populations. A large-sample structural MRI investigation of 217 young healthy adults tested whether hippocampal grey-matter volume could explain performance on multiple hippocampal-dependent tasks, including navigation, and reported little evidence for a robust link between volume and behavior across methods and subgroups. A separate analysis of roughly 99 healthy young adults found that the relationship between hippocampal measurements and navigation depends heavily on which subfields are measured and how sex differences are handled, with Simpson’s paradox complicating results that look straightforward on the surface. The hippocampus, a small structure deep in the brain that is central to memory and spatial orientation, clearly matters for navigation, but its gross volume appears to be a poor proxy for how well it actually performs that job in everyday life.
Volume Versus Connectivity: A Measurement Problem
A peer-reviewed synthesis of the mixed hippocampal volume literature, published in Neuroscience and Biobehavioral Reviews, offers several explanations for why structural findings keep falling short. The authors argue that the taxi driver results may capture extremes of experience rather than the typical range of human variation, making them a poor template for generalizing to most people. They also note that measurement choices (such as which segmentation protocol is used, whether total hippocampal volume or subfield volumes are analyzed, and how covariates like age and sex are handled) introduce substantial noise that can obscure or even invert true relationships between anatomy and behavior.
Most critically, the review raises the possibility that connectivity between brain regions, not the volume of any single region, may be the more relevant factor for navigation and related cognitive skills. Structural MRI excels at quantifying the size and shape of brain tissue, but it is relatively insensitive to how neurons are wired together or how activity flows through large-scale networks during complex tasks like wayfinding. The weak predictive performance of Weisberg’s AI models, combined with training studies that show functional reorganization without structural change, suggests that future work may need to prioritize measures of connectivity and dynamics, such as diffusion imaging, resting-state functional connectivity, and task-based activation patterns, over static snapshots of regional volume.
Rethinking What AI Should Learn From the Brain
Taken together, these converging results point toward a reframing of how neuroscientists use AI. Rather than expecting deep learning systems to wring behavioral predictions out of structural images alone, researchers may be better served by feeding these models richer, multimodal data that combine anatomy with functional and connectivity measures. In that context, AI could be used to uncover patterns in how networks reorganize during navigation training, or to identify latent dimensions of brain dynamics that align more closely with real-world wayfinding success than any single structural metric does.
For now, the failure of sophisticated algorithms to decode navigation skill from structural MRI scans serves as a cautionary tale about overinterpreting what brain volume can tell us. It underscores that even powerful computational tools cannot conjure predictive signals that are not present in the data to begin with, and it invites a broader shift in focus from static structure to flexible function. As labs continue to integrate AI into cognitive neuroscience, Weisberg’s work suggests that the most informative questions will not be about how big a brain region is, but about how its activity patterns and connections adapt when we learn to navigate the world around us.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
