Beyond Output Quality: Participation Quality as a Distinct Evaluation Construct and Control Objective for Human–AI Systems

This paper introduces Participation Quality (PQ) as a distinct evaluation construct for human–AI systems and Participation Optimization as the corresponding control problem. Contemporary AI evaluation is already multidimensional, addressing capabilities, output quality, safety, calibration, robustness, fairness, efficiency, trust, reliance, and human–AI collaboration. This paper identifies a complementary question that these dimensions do not fully capture: whether an AI system selected an appropriate way to participate in a human situation. Participation Quality evaluates the form, timing, degree, and authority of AI participation. A system may produce an accurate, useful, safe, and well-calibrated output while still participating poorly—for example, by answering when it should ask, deciding when judgment should remain with the human, intervening too early or too strongly, executing when approval should first be obtained, remaining silent when intervention is warranted, or failing to return control. We refer to such failures as participation errors: failures of participation selection that need not be failures of intelligence or output quality. The framework distinguishes Participation Process Quality from subsequent Human–Relational Effects and Downstream Outcomes, allowing participation fit to be evaluated without reducing it to immediate user satisfaction or outcome success. Participation Process Quality evaluates whether the AI selected an appropriate form, timing, degree, and authority of participation, while the latter layers examine the consequences of that participation for human agency, reliance, control, relationships, decisions, tasks, and longer-term interaction. The paper further formulates Participation Optimization as a constrained, multi-objective control problem in which intervention optimization occurs only after an appropriate participation configuration has been selected. Output and action quality therefore remain essential, but are treated as conditional on an upstream participation decision. The resulting architecture extends the conventional pattern: input → optimize output toward: context → governance → participation selection → optimized intervention → effects → updated context. The paper also proposes an empirical research program comparing identical foundation models under baseline, matched-orchestration, and Participation Governance conditions. This design is intended to test whether participation-aware control produces measurable Participation Lift beyond underlying model capability, output quality, and generic orchestration effects. The framework does not claim that abstention, clarification, mixed initiative, appropriate reliance, human agency, levels of automation, shared control, delegation, or human–AI collaboration are individually new. Rather, it proposes that these phenomena can be understood as related manifestations, constraints, consequences, or special cases of a more general system-level variable: AI participation. The central proposition is: High output quality does not imply high participation quality. This paper serves as the evaluation and optimization layer of the broader Participation Governance research program, extending the analysis of AI from what systems can produce toward whether, how, when, and with what authority they should participate in human situations.