ABSTRACT
Artificial intelligence (AI) has become a prominent public issue, particularly in China, where the government has announced plans to turn the country into a global AI power. This study analyses public discourse about AI in China through the conceptual lens of public spheres theory and counter-public spheres. It compares the official AI narrative on People’s Daily Online with public discussion about AI on the social medium WeChat, where we assumed that official views would be challenged. Using a combination of qualitative and computational methods, 140,000 AI-related articles published between 2015 and 2018 were studied. Findings reveal that AI-related discourse on WeChat is surprisingly similar to that on People’s Daily Online. That is, it is dominated by industry and political actors, such as government agencies and technology companies, and is mostly characterised by discussions about the economic potential of the technology, with strongly positive evaluations, and little critical debate.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Notes
1 WeChat articles published by public accounts have various styles. In most cases, they look like news articles or long blog posts.
2 This pre-processing step involved removing stop words, sparse terms, and dense terms. We trained LDA models with combinations of k (k = {10, 20, … , 80}) and hyper-parameter Alpha (alpha = {.005, .01, .05, .1, .2, .5, 1}). Beta was set at 1/k (Maier et al., Citation2018). For each k, we selected a model with the best topic coherence.
3 ‘Boilerplate’ topics are those reflecting general language-usage features.
4 Our LDA models could generate the topic membership probability θt of an article for all topics. These probabilities are mutually exclusive for each article, i.e. θt1 + θt2 + … + θtk = 1. Suppose we grouped t1, t2 and t9 into frame x (fx). The frame membership of an article to frame x (θfx) is equal to the sum of θt1, θt2 and θt9. According to the sum rule, θfx can be interpreted as the probability of an article falling into t1, t2 or t9.
5 In Epidemiology, RR is commonly calculated by the division of two risks. To translate this concept into our sentiment analysis, we calculated the RR by dividing the ratio of positive words by the ratio of negative words. Because the bases of the two ratios are the same, the RR value equals the number of positive words divided by the number of negative words. We added one to the number of negative words to solve the problem of division by zero.
6 Through an analysis of governmental policy documents, we developed a timeline of key AI-related policies launched in the past four years in China to interpret time series data in context (Appendix C).
7 Some sample institutions include The Institute for Ethical AI & Machine Learning (UK); The AI Now Institute (US); and The Future of Life Institute (US).
8 The is no mean theta value for PD’s socio-ethical frame because we did not identify any prevalent topic under this frame in our topic modelling analysis (see ).
Additional information
Notes on contributors
Jing Zeng
Dr. Jing Zeng is a senior research associate in the science communication division in the Department of Communication and Media Research (IKMZ), University of Zurich. Her research interests include science communication, digital culture, and online misinformation.
Chung-hong Chan
Dr. Chung-hong Chan is a research fellow at the Mannheimer Zentrum für Europäische Sozialforschung (MZES), University of Mannheim (Germany). An epidemiologist by training, he is interested in developing new quantitative methods for social science research.
Mike S. Schäfer
Dr. Mike S. Schäfer is professor of science communication and Director of the Center for Higher Education and Science Studies (CHESS) at the University of Zurich. He also heads the AGORA program of the Swiss National Science Foundation, and an expert group on the future of science communication at the Swiss Academies of Arts and Sciences. His research focuses on science communication, online communication, and science-related attitudes.