WHO HOLDS THE REINS? Facilitating Mixed Human-AI Teams: Goal, Execution, and Verification.

Abstract We propose a theoretical model for organising joint work between humans and large language models (LLMs) grounded in the theory of group facilitation. Experimental research on group work indicates that the best-supported mechanisms for improving the quality of group outcomes are structural. Applying this conclusion to LLMs, as a heuristic analogy with explicitly stated limits, leads to a distinction between facilitation, which operates through the logical parameters of a task, and manipulation, which operates through prompt features unrelated to the task. We show that, in existing studies, the intrinsic effect of emotional tone in prompts has not been separated from confounding factors (instruction clarity, prompt structure, social pressure) and replicates poorly; this does not mean that emotionally framed wording has no influence on the result. Mixed work is described by a three-layer architecture of goal, execution, and verification. Mapped onto it are a role matrix, Who Holds the Reins, defined by task type and relative execution competence, and the radius of AI autonomy. The central hypothesis of the article is that, when AI is stronger in execution, the boundary of applicability of the key configuration, in which the AI performs while the human sets the goal and verifies, is determined by the human's verification competence and the structure of verification rather than by execution competence. The hypothesis is motivated by scalable oversight research, which establishes this possibility in model tasks but has not tested it in mixed teams, and leads to the notion of a blind spot, which can be addressed by changing the structure of verification rather than by reallocating roles. We distinguish established findings, propositions, assumptions, and hypotheses, and formulate nine testable hypotheses together with a methodology for testing them.