AI systems lack a crucial safety mechanism that human bodies provide, UCLA researchers say
Artificial intelligence models like ChatGPT and Google's Gemini can describe human experiences with fluency but have no way to actually experience them. A new study published in Neuron argues this gap has real consequences for AI safety and alignment.
UCLA Health researchers propose that current AI systems are missing two essential ingredients: a body that interacts with the physical world and internal awareness of that body's states-fatigue, uncertainty, physiological need. The researchers call this combined property "internal embodiment."
The distinction matters beyond philosophy. In one test, the researchers showed several leading AI models a simple image of moving dots arranged to suggest a human figure-a test that even newborns recognize as human movement. Several models failed to identify the figure as a person. One described it as a constellation of stars. When the image rotated just 20 degrees, even the best-performing models broke down.
Humans don't fail this test because human perception is anchored to a lifetime of bodily experience. AI systems trained on vast libraries of text and images but with no bodily experience are pattern-matching without that anchor.
External versus internal embodiment
Current generative AI and LLM research focuses on "external embodiment"-a system's ability to interact with the physical world, perceive its environment, and respond to feedback. That work is important.
But "internal embodiment" has been largely ignored. This means continuously monitoring one's own internal states. Humans do this automatically through organs, hormones, and the nervous system. That information shapes attention, memory, emotion, and social behavior.
Current AI systems have no equivalent. They process inputs and generate outputs without any persistent internal state that regulates behavior over time. Without internal costs or constraints, an AI system has no intrinsic reason to avoid overconfident errors, resist manipulation, or behave consistently.
A framework for safer AI
The researchers propose a "dual-embodiment framework"-a set of principles for building AI systems that model both external interactions and internal states. These internal state variables would not replicate human biology but would function as persistent signals tracking uncertainty, processing load, and confidence.
The authors also call for a new class of benchmarks to measure internal embodiment. Existing AI tests focus almost exclusively on external performance: Can the system navigate a space? Identify an object? Complete a task?
The field needs evaluations that probe whether a system can monitor its own internal states, maintain stability when those states are disrupted, and behave in ways that emerge from shared internal representations rather than statistical mimicry.
Research in this area could shift how developers approach AI alignment. If companies want systems genuinely aligned with human behavior-not just superficially fluent-they may need to build in vulnerabilities and checks that function like internal self-regulators.
Your membership also unlocks:
