Bio Mirror / Memory, Motor Control, and Embodiment
Proprioception: The Sense That Makes Control Possible
Every controller needs a state estimate. The body has dedicated organs for producing one, and losing them is catastrophic.
The structure
Muscle spindles, embedded in the muscles themselves, signal length and rate of change of length. Golgi tendon organs signal force. Joint receptors and cutaneous mechanoreceptors contribute position and contact information. Crucially, spindles have their own motor innervation — gamma motor neurons set the spindle's sensitivity, so the nervous system actively tunes its sensors according to what it expects to happen.
The function it dictates
Proprioception supplies the state estimate that every internal model requires. Without it, movement is possible but nearly unusable: the rare patients who have lost proprioception through sensory neuropathy can move only by watching their limbs continuously, and cannot walk in the dark. The system also integrates proprioception with vision and the vestibular sense, weighting each by reliability — which is the same computation a sensor-fusion filter performs, and it produces the same characteristic illusions when one channel is manipulated.
The mirror in AI
The nearest analogue is a model's access to its own state. This is thinner than the other mirrors in this section, and the thinness is the point: a language model has no proprioceptive channel. It cannot sense its own uncertainty as a signal in the way a limb senses its own position, and the calibration failures in the DEM-X catalog are downstream of that absence. A model has no organ for knowing what it does not know.
The mirror in robotics
Joint encoders, force-torque sensors, and IMUs feeding a state estimator. Every robot has this and every roboticist knows the system is only as good as its state estimate. The gamma motor system has a striking engineering parallel in adaptive sensor gain — adjusting a sensor's sensitivity range based on the expected operating condition — though few robots implement it as aggressively as biology does.
Biological proprioception is predictive rather than merely reportive: the nervous system issues an efference copy of each motor command and compares the predicted sensory consequence against the actual one. That comparison is how you distinguish self-generated from external sensation, and its failure is implicated in the experience of alien control in schizophrenia. No deployed AI system has an equivalent mechanism for distinguishing the consequences of its own actions from changes in the world.
Open questions
- What would a genuine proprioceptive channel for a language model look like?
- Is the efference-copy comparison a general principle that artificial agents should implement?
Related DEM-X entries
INF-HALL-01Further reading
- Proske & Gandevia, "The proprioceptive senses" (2012)
- Cole, Pride and a Daily Marathon — a first-person account of proprioceptive loss