Human-AI Collaboration Needs Synchronization

AI systems are evaluated, trained, and deployed as if they were solo performers, even when humans collaborate with them. Yet every deployment places a machine operating at thousands of tokens per second alongside a human processing roughly ten bits per second, across a communication channel with limited pacing, weak shared-state mechanisms, and limited means for either party to verify the other’s understanding. This position paper argues that the resulting failures, cognitive overload on the human side, silent drift on the AI side, and an accountability gap that neither party can close, are not only failures of model capability but also failures of an undesigned channel. We contrast the human-to-AI channel with the human-to-human and AI-to-AI channels, show that it lacks many of the robust coordination mechanisms available in the other two channels, and document how it breaks down from both ends. We argue that the channel itself must become a first-class research object, with evaluation, training, and interaction protocols engineered specifically for it. Without intervention at the channel level, improvements in model capability will continue to widen the gap between what AI produces and what humans can meaningfully account for.