摘要
AI is rapidly becoming embedded in design practice, from individual ideation to team collaboration and collective decision-making, yet evaluation still privileges short-term output metrics. We argue that outcome-centered paradigms are inadequate for Human-Centered AI because they under-measure the experiential conditions that sustain design innovation over time, including agency, creative self-efficacy, ownership, and calibrated trust. Drawing on HCI and creativity psychology, we conceptualize design innovation as a dynamic, situated process shaped by autonomy and competence, and by interactional work of synthesis, trade-offs, and attribution. As LLMs fluently generate concepts, proposals, and summaries, designers may shift from originators to selectors; even when outputs appear strong, this shift can weaken perceived ownership and diminish creative self-efficacy. At the team level, AI mediation can redistribute conversational authority, obscure provenance, and compress divergence, intensifying convergence pressures and participation inequities. For small-societies, legitimacy becomes central: participants must understand and contest how inputs are represented in the evolving collective record. We propose a design approach that treats LLM-based and multi-agent systems as configurable socio-technical mediators, articulated through seven principles: role-differentiated mediation, legibility of influence, provenance, contestability, divergence before convergence, reflection scaffolds, and trust calibration. We conclude with an evaluation agenda that complements output metrics with longitudinal measures of agency, innovation self-efficacy, ownership, participation equity, and procedural legitimacy.
