Humans have an innate ability to perceive possible actions in their environment—knowing, for instance, that a path can be walked on or that a lake can be swum in. Despite recent advances, artificial intelligence (AI) still struggles to match this level of intuitive understanding. A new study from the University of Amsterdam sheds light on the unique brain activation patterns that underpin how people navigate physical spaces, offering insights that could help bridge this gap in AI.
Using MRI scans, researchers observed brain activity while participants viewed images of indoor and outdoor environments and judged appropriate actions—such as walking, cycling, or swimming. The study found that specific regions of the visual cortex were activated in ways that not only reflected the visual content but also automatically processed possible actions. This phenomenon, known in psychology as affordances, occurs even without conscious deliberation: the brain intuitively assesses what actions are possible within a given setting.
The team also compared human performance with that of AI systems, including image recognition models and GPT-4. The findings revealed that AI systems performed poorly when predicting potential actions in a scene. Even with task-specific training, their internal computations did not resemble human brain activity. This highlights a fundamental difference: human visual understanding is deeply rooted in real-world physical experience—something AI currently lacks.
The study has far-reaching implications for AI development, particularly in fields like healthcare, robotics, and autonomous vehicles, where systems must not only recognize objects but understand how they can be used or interacted with—for example, helping robots navigate disaster zones or enabling self-driving cars to distinguish bike lanes from roads. As AI training grows increasingly energy-intensive and centralized, mimicking the brain's efficiency could lead to smarter, more sustainable, and human-aligned AI systems.