Why do researchers refuse to delegate tasks to AI agents?

The formative study behind Cocoa (9 researchers) found participants "preferred to perform higher-level reasoning and synthesis themselves" and were "often unsatisfied with AI's outputs on these tasks" — but the paper's discussion separates two distinct reasons for withholding a task from an agent, only one of which is about output quality. P3 said "this tinkering process around reading and playing around with things is what gives you the ideas. I don't know if I want those things automated because the process is as helpful as the final result," and P5 wanted the agent to merely "give me inspiration for where I can go" rather than "making the final decisions."

The discussion section names the two reasons explicitly: "First, the user can perform the task more effectively than the agent," typically because "the user holds more expertise" (their example: brainstorming research questions in one's own research area); "Second, the task is intrinsically valuable for the user to do themselves, even if an agent is capable of automating it" (their example: "reading seminal papers to develop research taste"). The authors draw a design consequence from the split: systems for expert users who want to keep exercising judgment should offer many levels of control throughout planning and execution, while systems for users navigating unfamiliar domains without an upskilling goal can offer fewer controls, since "determining the user's capacity and willingness to take on more cognitive load can inform the appropriate level of control to afford."

This complicates the assumption, implicit in Do humans and AI divide labor differently in knowledge work?, that collaboration mode is mainly a function of who is better at deciding task and method; Cocoa's formative study adds a motive for retaining a task that has nothing to do with comparative skill at all — the process of doing the task is what builds the human's own future capability, so a user might keep a task even while rating the agent's own output as adequate. It also qualifies How should AI agents and humans divide research tasks?: that record of who makes final decisions in practice does not distinguish whether humans retained decisions because they were better at them or because making them was itself worth doing. The system this finding motivated is described in Can users steer AI agents better by revising plans mid-execution?.

The excerpt gives this split as a stated reason drawn from 9 formative-study participants, not as a measured comparison of task outcomes under the two motives, so the paper does not establish how often "expertise" versus "upskilling" is the operative reason in a given refusal to delegate, nor whether either reason predicts behavior outside scientific research. The implication the authors draw, and the excerpt supports only this far, is that interaction designs should expose control over a task independent of how well an agent can already perform it — because a well-performing agent is not sufficient grounds to assume a user wants the task delegated.