Maintaining Sufficient Interactions for Accountable AI: An Executable Transparency Audit Framework in Recommender Systems

AI-driven recommender systems have deeply integrated into digital life. However, their "black box" nature and associated societal risks profound transparency issues. Existing transparency solutions often fail to engage multi stakeholders due to the dynamic, non-deterministic characteristics of AI models. To address this challenge, we reconceptualizes transparency not as static information disclosure, but as a dynamic interactional process within a socio-technical network through introducing the Actor Network Theory. Grounded in this theoretical lens, we propose an executable transparency audit framework. By employing Natural Language Processing techniques, we systematically analyze regulatory texts and expert interviews, thereby constructing a transparency assessment checklist. Applying this framework, we conduct an empirical validation on Chinese social media platforms. The results reveal the pervasive phenomenon of "paper compliance", a strategy where platforms adhere to regulative mandates through formalized compliance while neglecting the intrinsic intent. Further analysis delineates two distinct patterns of this symbolic compliance: "formal regulation-oriented" and "selectively interactive-oriented". In-depth experiment identifies insufficient interaction among platforms, users, and regulators as the root cause underpinning the lack of substantive transparency. This study enhances the concept of transparency from an interaction based perspective and developing an operational transparency audit framework. Moreover, the paper uncovers the manifestations and formation mechanisms of "paper compliance", offering a critical foundation for optimizing regulatory policies and incentivizing substantive governance.