Measuring Human Labor in Human-Agent Joint Production

In human-agent workflows, a single user instruction can redirect an agent's search, reasoning, and tool use. Yet the final output typically gives no indication of which instruction mattered or how much human effort went into shaping it. A new arXiv preprint argues that measuring human labor requires shifting the unit of analysis from the whole task to the individual turn.

The paper, "From Tasks to Turns: Measuring Human Labor Contributions in Human-Agent Joint Production," proposes that contribution should be assessed at the level of each instruction or turn, rather than inferred from the end result. This matters because the same task outcome could be achieved with very different amounts of human guidance, and current evaluation methods obscure that difference.

The abstract is brief and does not detail the proposed measurement method, but the framing suggests a move toward more granular accounting of human effort in AI-assisted work. For researchers studying human-AI collaboration, this could provide a way to compare workflows and understand where human input actually changes outcomes.