The next speaker in this session at the 2026 International Communication Association conference in Cape Town is Yingdan Lu, whose focus is on the impacts of social media algorithms on the curation of state-created content in China. Authoritarian governments are of course increasingly leveraging algorithmic systems for their digital propaganda; this both censors critical information, promotes pro-regime materials, and floods social media spaces with politically irrelevant content in order to make critical content less easy to find.
The focus here is on recommendation algorithms, and explores algorithmic promotional curation processes which systematically amplify state-created content. In China, social media platforms are under strict state control, and by law algorithmic recommendations must adhere to public values and state priorities; it is assumed that this promotes state content, but there is limited empirical evidence to date that this is the case. But state algorithmic control might manifest in more subtle ways, of course.
The focus here is on Bilibili, the top video-sharing platform in China, which has a highly curation-dependent platform design with a trending video page and personalised content feeds, and is dominated by entertainment content. It is also under strict state control, however, and is collaborating with the state to promote propaganda content.
The project collected the 100 trending videos on each our, for 91 days; this also included the top 10 videos for each recommended video. It distinguished between state and non-state accounts, and matched 195 state and 193 non-state accounts for which it also collected all their Bilibili homepage videos. The videos were then labelled from the metadata, and categorised.
For each trending video from state accounts, over 80% of the recommended further videos were also by state accounts. For each homepage video, the percentage of further recommended videos was also very high for state accounts, compared to non-state accounts. The state affiliation of trending videos consistently predicts a high rate of further state video recommendations, but this is not uniform across all content categories; it is still high, but lower for entertainment and other content categories than it is for political content.
Simulating this over many iterations, this produces a certain mix of video types: even after iterating through multiple steps, recommended content following state news and politics content will still be highly state news and politics-dominated; for state entertainment content it declines more quickly. For non-state videos, the mix diversifies more quickly.
State content is thus systematically reinforced through algorithmic promotional curation, but this varies by content categories. Algorithmic promotional curation might thus balance state and commercial interests here; the algorithm serves authoritarian goals, but also satisfies the company interests. Such patterns are likely to be common in other authoritarian regimes beyond China, too.
