The burgeoning of generative artificial intelligence is reshaping the paradigm of scientific knowledge dissemination，while its derivative phenomenon of “knowledge hallucination” profoundly deconstructs the foundational credibility of scientific communication. Drawing upon Latour’s actornetwork theory analytical framework，this paper posits AI knowledge hallucination as a product of systemic distortion or “betrayal” within the translation process. This occurs within a heterogeneous actornetwork due to the divergence of objectives and conflicts of interest among key actors. This phenomenon stems from four structural dysfunctions：Source networks suffer contamination from diachronic biases in data inscription and algorithmic goal alienation；Error-correction networks face institutional feedback deficits and temporal asynchrony conflicts hindering rectification; Translation chains endure professional discourse dimensionality loss and technical violence from cross-contamination within heterogeneous knowledge networks；Accountability networks descend into governance vacuums due to technological black-box obscurity and fragmented responsibility. To address these issues，this paper proposes a four-dimensional path for collaborative governance：“Source Purification” fortifies data foundations through a multi-centre knowledge certification system and blockchain traceability technology；“Process Streamlining” builds a closed-loop error correction network leveraging intelligent monitoring systems and dynamic knowledge repositorie；“Translation Optimisation” ensures knowledge fidelity via context-aware algorithms and human-machine dual verification mechanisms；“Anchoring Accountability” clarifies responsibility boundaries among diverse stakeholders through legal empowerment and transparent algorithmic disclosure. This research deepens understanding of the socio-technical nature of AI knowledge hallucination，advancing the development of a robust，trustworthy “AI+Science Communication” ecosystem centred on human-machine collaboration.