Authoritarian Innovation: How China Is Building Control into Generative AI Systems

(Scroll Down for Traditional Chinese version)

Written by Resilience Innovation Lab (RIL) and Strategic China Policy Group (SCPG), the latest report, Authoritarian Innovation: How China Is Building Control into Generative AI Systems, draws on official sources and technical standards to examine how Beijing is embedding information control into large language models through its content security framework, standards, and model governance.

Key findings:

1. China treats content that departs from Party narratives as a “content security risk” — built into AI systems from the ground up
2. An expanding system of industrial standards now governs training data, model behaviour, and content testing across the entire AI pipeline
3. Private firms like Alibaba and DeepSeek aren’t just regulated — they’re incorporated into the governance system itself
4. As Chinese models spread abroad (7 of the world’s top 10 LLMs by usage are now Chinese), this control logic travels with them

Read the report here.

威權創新：中國如何將管控植入生成式人工智能系統

韌性創新實驗室（Resilience Innovation Lab, RIL）與Strategic China Policy Group （SCPG）合著的最新報告《威權創新：中國如何將管控植入生成式人工智能系統》（Authoritarian Innovation: How China Is Building Control into Generative AI Systems），透過官方資料及技術標準，探討北京如何透過其內容安全框架、行業標準及模型治理機制，將資訊管控嵌入大型語言模型之中。

報告重點：

1. 中國將偏離黨的敘事的內容視為「內容安全」風險，並自源頭起將管控嵌入人工智能系統
2. 一套不斷擴張的行業標準體系，正涵蓋訓練數據、模型行為及內容測試等生成式人工智能全流程
3. 阿里巴巴、DeepSeek 等私營企業不僅受監管，更被納入治理體系本身
4. 隨著中國模型全球擴散（全球十大主流大型語言模型中已有七個來自中國），這套管控邏輯亦隨之外溢

請按此閱讀全文（只有英文版本）。