Projecting Frames About AI: Chinese State Media Between the Global South and the Global North

Chinese state-affiliated international media have become a consequential presence in the global information environment, and artificial intelligence has become one of their most contested subjects. AI stands simultaneously as a frontier of international norm-making, a pillar of Chinese industrial policy, and an arena of Beijing–Washington competition. Whether Beijing's outward-facing media address international audiences uniformly on this subject, or calibrate their coverage to different regions, has remained unexamined at scale. This paper asks two questions. First, do frames in Chinese state-media AI coverage, understood as selective emphases on aspects of the same issue rather than distinct topics, vary systematically between content addressed to Global South and Global North audiences? Second, does this coverage track the language of China's codified AI regulations equally across audiences? The study compares three bodies of text on a shared set of inductively identified frames: Xi Jinping's speeches on AI, from which eight frames are derived, a corpus of China's codified AI regulations, and 316,881 items collected from twenty-five outward-facing state-media channels on YouTube and TikTok between January 2022 and March 2026. Three large language models of Chinese, US, and European provenance score these corpora in parallel under a sealed configuration, a tri-provenance design that doubles as a test of model provenance bias. Hypothesis tests use channel-respecting inference throughout. Both questions return a differentiated picture. Frame emphasis varied systematically by audience (χ²(7) = 262.86, Cramér's V = 0.29, n = 3,127). Coverage addressed to the Global North was led by AI ethics and safety, at nearly four times its Global South rate, together with economic growth and technological self-reliance. Global South coverage foregrounded global cooperation and shared benefits, public-service applications, talent cultivation, and integration into the real economy, while scientific achievement travelled to both audiences at near-identical rates. Regulatory alignment was modestly but consistently higher for Global South audiences on both reference texts (Cohen's d = 0.31 and 0.32), a tilt registered by all three models and corroborated by an independent embedding measure. Global South content was however excluded from the alignment comparison at markedly higher rates, chiefly because short clips carried no regulatory language. The findings suggest that Chinese state media practise multivocal signalling through frames: emphasis is segmented by audience while the underlying regulatory vocabulary remains broadly common, the configuration least exposed to cross-audience comparison on open platforms. The paper contributes the first large-scale evidence of audience-differentiated AI framing across the outward-facing state-media system, extending work on authoritarian norm entrepreneurship and norm subsidiarity from official documents and diplomacy to mass communication. Methodologically, it presents a portable tri-provenance ensemble protocol for inductive frame analysis without an expert-coded benchmark, of direct relevance to research on algorithmic information control and state communication on platform media.