Collaboration Bottleneck
Narrow channel between human and AI that limits how much context reaches the model.
The collaboration bottleneck is the fundamental limitation that prevents fluid human-AI teamwork: today's AI models experience interaction in a single sequential thread, unable to perceive what a user is doing or thinking while they wait. Until the user finishes typing or speaking, the model has no awareness of context, intent, or nuance being communicated in parallel. Until the model finishes generating, its perception freezes — receiving no new information and offering no opportunity for course correction, clarification, or interruption. This creates a narrow, rigid channel for human-AI collaboration that arbitrarily limits how much of a person's knowledge, intent, and judgment can reach the model, and how much of the model's reasoning can be understood before it becomes irrelevant.
The bottleneck is structural, not technological. It emerges from treating AI interaction as a series of discrete turns rather than a continuous two-way exchange. Turn-based interfaces force humans to contort their natural communication patterns — which are rich, overlapping, and context-dependent — into a rigid send-then-wait pattern. This is analogous to resolving a complex disagreement over email rather than in person: the medium strips away the immediacy, nuance, and responsiveness that make collaboration effective. The result is that humans increasingly get pushed out not because the work doesn't need them, but because the interface has no room for them.
The consequence is that humans can neither correct AI errors in real time nor leverage the AI's work as it emerges. A user who notices an AI going down the wrong path midway through generation must wait for completion before intervening — at which point significant wasted work may have occurred. This is especially costly in complex, high-stakes tasks where the value lies in iterative refinement through dialogue rather than one-shot generation. Research from METR (2025) and frontier model evaluations consistently find that autonomous, long-running agent harnesses outperform interactive synchronous patterns — a signal that today's interfaces are misaligned with how humans actually want to work.
The bottleneck can only be resolved by making AI interactive in real time across any modality, enabling AI interfaces to meet humans where they are rather than forcing humans to contort themselves to AI interfaces. This is the core motivation behind interaction models, full-duplex audio systems, and other continuous-interaction architectures.