UCLA Study Reveals 'Body Gap' in AI: Language Models Can Describe Human Experience But Lack Embodied Understanding
Key Takeaways
- ▸Current AI systems can describe human experiences linguistically and visually but cannot truly understand them through embodied experience, creating a fundamental gap in their understanding of the world
- ▸The lack of 'internal embodiment'—continuous monitoring of internal states like fatigue, uncertainty, and physiological need—results in measurable performance failures, such as difficulty recognizing human motion patterns that newborns can identify
- ▸Implementing functional analogies of embodied experience in AI could be crucial for improving both performance and safety, particularly for AI deployed in high-stakes domains
Summary
A new study published in Neuron by UCLA Health researchers argues that today's most advanced AI systems, including multimodal large language models like ChatGPT and Google Gemini, lack a critical component of human intelligence: embodied experience. The researchers propose that current AI systems are missing two essential ingredients—a body that interacts with the physical world and an internal awareness of that body's states such as fatigue, uncertainty, or physiological need—which they collectively term "internal embodiment."
The study highlights how this "body gap" has measurable consequences for AI performance and behavior. In one experiment, several leading AI models failed to recognize a point-light display (dots arranged to suggest a human figure in motion) that even newborns can identify, with some describing it as a constellation of stars. When the image was rotated just 20 degrees, even the best-performing models broke down. Humans excel at this task because their perception is anchored to a lifetime of bodily experience, whereas AI systems trained on vast text and image libraries engage in pattern-matching without this fundamental grounding.
The researchers distinguish between "external embodiment"—a system's ability to interact with the physical world—which is already a focus in current multimodal AI development, and "internal embodiment," which has not been implemented in these models. According to the authors, this internal dimension acts as a built-in safety system in humans, allowing them to register uncertainty, depletion, or threats to survival. The absence of such mechanisms in AI poses significant implications for safety and trustworthiness, particularly as these systems are deployed in consequential real-world settings.
- The distinction between external embodiment (interaction with the environment) and internal embodiment (awareness of one's own states) represents an underexplored but critical frontier in AI research
Editorial Opinion
This UCLA study identifies a fundamental limitation in how we've architected modern AI systems that goes beyond typical capability gaps. The insight that AI systems lack the embodied anchor that humans rely on for perception and decision-making is philosophically important and practically consequential—it suggests that scaling up data and parameters alone won't solve certain classes of problems. If the researchers are correct that internal embodiment acts as a built-in safety mechanism in humans, then addressing this gap isn't merely an academic exercise but a potential prerequisite for deploying AI systems responsibly in high-stakes environments.
