[Purpose/significance] As generative artificial intelligence (GAI) grows more dependent on training data,super platforms have used their structural advantages to dominate model training, creating a "platform-model" symbiot⁃ic relationship. However, existing research has paid little attention to the operational mechanisms of platform control and the associated governance dilemmas within the training data phase. [Method/process] Using normative analysis and institutional comparison, this paper dissects the unique power structure and risk types inherent to the "platformmodel" symbiosis and reveals their risk formation and governance pathways. [Result/conclusion] The study finds that training data risks under this symbiotic relationship are endogenous and amplified, challenging traditional regulatory models. To address this, it proposes a multi- stakeholder collaborative governance system featuring: a principled,tiered, government-led source documentation framework to manage privacy leakage and data crossover risks; a verifi⁃able platform self-assessment system to dismantle governance barriers; and superplatform-led construction of trusted data spaces to mitigate data circulation risks.