A new arXiv paper tries to explain why human-AI teams keep outperforming either one working solo, and names the shapes that teamwork actually takes.
The paper examines two design dimensions that shape any human-AI collaboration: how much autonomy the AI gets, and who takes the initiative. Rather than just listing design tips, the authors use a "paradox" lens, tracing tensions back to root problems, then mapping how those tensions get resolved in practice. From that process they derive four named patterns: Instruction, Delegation, Assistance, and Co-creation.
Most AI product debates boil down to a vague instinct about how much control to hand over. This paper is useful because it turns that instinct into named, comparable categories instead of a one-dimensional slider from "human in charge" to "AI in charge." Instruction and Delegation sit at opposite ends of who initiates the work, while Assistance and Co-creation describe different depths of back-and-forth once both parties are involved.
None of this is validated against real products yet - it is a conceptual framework, not a user study. Whether these four patterns hold up once someone tests them on actual AI copilots and agents is the obvious next question.