UCLA researchers are throwing some serious cold water on the hype around advanced artificial intelligence, arguing that today’s most powerful models are missing a basic ingredient humans take for granted: any real internal sense of their own bodies and states. Without that, they warn, systems can sound fluent and even insightful while staying fundamentally brittle and potentially unsafe in real-world settings.
The team points out that language-and-vision models can talk about pain, fatigue or doubt, but they do not have persistent internal signals that correspond to those experiences. There is no built-in feeling of being tired, overloaded or uncertain to help regulate their behavior. According to the researchers, those blind spots already show up on surprisingly simple perception tests and should give pause to anyone looking to deploy multimodal AI in medicine, transportation or public services.
New paper lays out the problem
The argument comes from a paper titled “Embodiment in multimodal large language models,” published in the journal Neuron. The work is summarized in a distribution on EurekAlert, where the authors describe what they call a lack of “internal embodiment” in current systems.
In plain terms, they are not just talking about robots without arms and legs. The critique is that modern multimodal LLMs do not maintain ongoing internal signals about their own condition. There is no continuously tracked sense of processing load, no intrinsic record of how confident they are, and no internal drive that nudges them to slow down or change course when something feels off. That, the team says, is a very different situation from the human brain these systems are sometimes compared to.
Point-light test exposes the gap
To show how this plays out in practice, the researchers turned to a classic psychology experiment: the point-light display. In this setup, a person’s joints are represented only by a handful of moving dots against a dark background. Humans, even without training, tend to recognize the pattern instantly as a person walking or moving.
Several leading multimodal models did not. When the team fed them point-light displays, multiple systems failed to identify the figure as human at all. In some runs, a model confidently labeled the pattern a “constellation” instead of a person in motion. That misfire, documented in the paper and covered by the Santa Monica Daily Press, may sound almost comical until you imagine the same kind of error in a clinical or safety-critical environment.
Proposed fix: dual-embodiment
The authors are not just diagnosing problems. They propose what they call a “dual-embodiment” framework for future AI. In their view, safer systems will need to model both the external world they interact with and an internal layer of ongoing states such as processing load, confidence and uncertainty.
On this view, an AI that is unsure of what it sees should be able to know it is unsure, represent that uncertainty internally and behave differently as a result. The team argues that the field also needs new benchmarks that explicitly test whether a system can monitor and stabilize its own internal variables, rather than just produce correct outputs on static tasks. Those recommendations are highlighted in coverage from UCLA Health.
Safety implications
Senior author Marco Iacoboni does not mince words about the stakes. “Without internal costs or constraints, an AI system has no intrinsic reason to avoid overconfident errors, resist manipulation or behave consistently,” he says, as quoted by UCLA Health.
The paper argues that giving AI some form of internal checks could approximate the self-regulation humans rely on every day. Our own bodies continuously feed the brain signals about fatigue, stress and risk, nudging us away from dangerous choices. Absent something analogous, the authors warn, advanced systems could make hazardous mistakes in clinical decision-making, transportation and civic contexts while sounding perfectly confident.
Why engineers and regulators should pay attention
For AI developers, the message is not exactly comforting. Building in detailed internal state models would likely mean giving up some architectural simplicity in exchange for systems that can self-monitor, flag uncertainty and decline tasks when conditions are not right. That could slow commercial rollouts, but the authors suggest it might also head off the kind of rare but dramatic failures that keep regulators and the public on edge.
The project is explicitly cross-disciplinary, blending perspectives from neuroscientists and AI practitioners at UCLA, USC and DeepMind. Independent coverage has underscored the paper’s call for safety-oriented benchmarks rather than just bigger and more capable models, as noted by Neuroscience News.
The authors are clear that they are not handing engineers a ready-made blueprint. Instead, they sketch out testable next steps, such as simulated interoceptive signals and benchmarks that probe how stable a system remains when its internal variables are perturbed. Full methods, datasets and suggested tests are detailed in the preprint hosted on arXiv, which lays out the technical backbone behind their warnings.
