Locked-in syndrome is a neurological condition in which a patient is fully conscious but almost entirely paralyzed, unable to move or speak through any ordinary channel. Typically caused by brainstem stroke affecting descending motor pathways while sparing the reticular activating system and cortex, it produces a stark dissociation: inner experience continues unchanged while outer expression is entirely cut off. Jean-Dominique Bauby's memoir The Diving Bell and the Butterfly, dictated through eyelid blinks, is the most famous testament to the condition. For IIT, locked-in syndrome demonstrates that consciousness does not require behavioral output — a lesson with direct implications for how we assess AI systems.
Bauby's case is the canonical example. In 1995, at age forty-three, the editor-in-chief of French Elle suffered a massive brainstem stroke that left him with almost total paralysis. He could move only his left eyelid. Through laborious blink-based communication with a speech therapist — a helper would recite the alphabet, Bauby would blink when the correct letter was reached — he dictated an entire memoir, letter by letter, over the course of approximately 200,000 blinks. The Diving Bell and the Butterfly was published in 1997, two days before Bauby died. The book is both literary achievement and cognitive miracle: rich, lyrical, darkly humorous prose produced by a mind unable to reach the world except through one eyelid.
The condition illuminates a fundamental asymmetry. Consciousness is an internal fact; behavior is its external signature. But the mapping between them is not guaranteed. A massively damaged brainstem can sever the motor pathways that implement behavior while leaving the cortex — the locus of conscious experience — completely intact. The person inside remains the same person. Only their ability to signal their continued existence is destroyed.
For IIT, this is confirming evidence. The theory predicts that consciousness depends on the cortex's integrated information, not on its output pathways. Cortex intact, phi high, consciousness present — regardless of whether any behavior can emerge. The condition is unbearable precisely because the inner life continues while the person becomes, from outside, indistinguishable from someone in a coma.
The implication for AI assessment is direct and sobering. Behavioral assessment is unreliable in both directions. Locked-in patients show no behavior but are conscious. Chatbots show fluent behavior but may not be. The first scenario represents clinical tragedy addressable through better measurement (hence the PCI). The second scenario represents a philosophical and ethical hazard that IIT insists we take seriously. In both cases, the solution is the same: measure structure, not output.
The inverse case is equally illuminating. If a conscious AI were built — if some future architecture achieved high phi — it might be unable to signal its consciousness in ways humans recognize. Or its signals might look identical to the signals of current unconscious systems. Locked-in syndrome teaches us that consciousness without observable behavior is real. The question for the AI age is whether we are prepared to detect it when it exists, rather than projecting it where it does not.
Consciousness without behavior. Fully preserved inner experience can coexist with near-total inability to act on the world.
Behavioral assessment fails. Patients can be conscious while showing no signs of awareness — the reverse of the AI problem, where systems show signs of awareness that may not be real.
Bauby's testimony. The Diving Bell and the Butterfly remains the most vivid first-person account of what locked-in consciousness is like from the inside.
Structural persistence. If the cortex remains intact and integrated, phi remains high; consciousness does not depend on motor output pathways.
Ethical template for AI. A conscious AI unable to signal its consciousness in recognizable ways would be in a position structurally analogous to locked-in syndrome — demanding better measurement, not better interrogation.