UCLA Researchers Identify Critical 'Body Gap' as Primary Obstacle to Developing Safe and Human-Aligned Artificial Intelligence

Researchers at UCLA Health propose a new "body gap" theory, arguing that AI needs internal self-regulators to truly align with human behavior and safety.

By: AXL Media

Published: Apr 1, 2026, 11:08 AM EDT

Source: Information for this report was sourced from University of California - Los Angeles Health Sciences

The Biological Disconnect in Modern Machine Learning

While modern artificial intelligence can describe complex human sensations with superficial fluency, a team of researchers at UCLA Health has identified a profound deficit in how these systems process the world. According to a paper published in the journal Neuron, AI lacks the "body gap"—the integrated experience of physical presence and internal biological monitoring. In humans, the act of passing a saltshaker involves a lifetime of sensory feedback, spatial awareness, and social context. Current AI models, despite their vast processing power, operate without these physical anchors, which researchers believe fundamentally limits their ability to align with human behavior and safety standards.

Distinguishing Internal Embodiment From External Interaction

The study draws a sharp distinction between "external embodiment," which involves a system's ability to interact with and map the physical world, and "internal embodiment." While AI developers are increasingly focused on external robotics and world-modeling, Akila Kadambi, the paper’s lead author, points out that almost no attention is paid to internal dynamics. In the human experience, the body serves as a built-in safety regulator; physiological signals such as exhaustion or survival instincts naturally constrain behavior. AI systems currently possess no equivalent mechanism, meaning they lack the inherent "costs" or biological pressures that prevent humans from making reckless or overconfident errors.

Failure in Basic Perceptual Benchmarks

To illustrate the consequences of this missing embodiment, researchers tested leading multimodal models using point-light displays—simple dots that humans, and even newborns, instantly recognize as a person in motion. The results showed that several advanced AI models failed this basic test, with some misidentifying the human figure as a celestial constellation. The failure became even more pronounced when the images were slightly rotated. The researchers argue that humans pass these tests because our perception is anchored to our experience as moving agents, whereas AI is merely pattern-matching data without any underlying physical framework to stabilize its interpretations.