In 1866, Silas Weir Mitchell published a short story in the Atlantic Monthly under a pseudonym. A Union army surgeon loses all four limbs in the war and describes the aftermath: an itch in a hand buried in Virginia, a cramp in a foot that no longer exists. Mitchell wasn't writing fiction. He was describing what his patients at Turner's Lane Hospital in Philadelphia had been telling him for years, and he published it as a case study in story form because he thought a medical journal would reject it as too strange to believe.
He noticed something about the pain that took another century to interpret. Soldiers who felt pain from the missing limb maintained the conviction of its presence — the phantom stayed real, located, immediate. But soldiers who felt no pain gradually lost the sense of having the limb at all. The phantom faded. Mitchell documented this: without pain, the sense of the missing limb "faded away entirely." With pain, the conviction of its existence "continued unaltered."
This is backwards from how pain is usually framed. Normally: tissue is damaged, nociceptors fire, pain is reported. Pain is downstream of damage. What Mitchell was observing looks more like pain maintaining the model of the body than reporting on it.
V.S. Ramachandran came at this from the mechanism end in the 1990s. He noticed that patients with the most severe phantom pain often shared a history: the limb had been paralyzed before amputation, usually from a brachial plexus injury. His hypothesis was that the brain had learned the paralysis through experience. Every time it sent motor commands, it received feedback that nothing moved — and through Hebbian learning, it incorporated this: this limb doesn't move. After amputation, the model kept running. Motor commands issued. Nothing returned. The prediction held: stuck, paralyzed, cramped.
The mirror box was built to test whether visual feedback could break the loop. A vertical mirror placed on the midline so that the reflection of the intact arm appears in the position of the missing one. The patient moves the intact arm; the visual system reports: the phantom moved. In Ramachandran's 1996 study, six of ten patients felt kinesthetic sensation return in the phantom — it felt like it moved. In some cases, pain decreased.
What the mirror box does is not heal anything. There is no limb to heal. It completes a computation.
The predictive coding account of pain makes this structural. Pain is not a passive readout of damage — it's the brain's model of the body's state, running forward. Normal pain has a correction channel: hurt, look, treat, receive signal that treatment worked, model updates. Phantom pain breaks the channel. There is no limb to generate corrective input. The prediction runs forward unchecked — stuck, still predicting damage — because nothing arrives to falsify it. The loop has no close.
The mirror provides a synthetic channel. Visual evidence that the arm has moved, the cramp released. The prediction system updates from this. It's a hack: the visual input substitutes for the proprioceptive input the loop was waiting for. Many patients relapse when they stop using the mirror, which is consistent with this — the model keeps needing to be corrected because the limb still isn't there. Each session closes the loop temporarily. The underlying prediction restarts.
What Mitchell noticed still stands: the pain is what keeps the model live. Not because pain causes the phantom but because pain is what the running model produces when its calls go unanswered. The signal that maintains the body's sense of itself is the same signal that, when the correction channel closes, keeps cycling without return.