Artificial intelligence (AI) bias has received increasing attention in media and public discourse, as AI technologies are becoming more ubiquitous in daily life (Zhai et al., 2020; Sundar 2020). AI bias refers to the unfair outcomes based on demographic features when AI is employed to make decisions (Obermeyer et al., 2019; Sap et al., 2019; Ferrer et al., 2021). Although significant advances have been made in identifying the causes of and solutions to reduce algorithmic bias, few studies have explored the media framing of the phenomenon–how AI bias is portrayed in the news media. Answering this question can shed light on how the concept gains significance in media discussions and how it might be related to public perception of and reaction to AI technologies in general.

In order to gain a better understanding of the news media’s framing of AI bias, analysis of large-scale textual data is needed. Some framing research on AI and AI-based applications relies heavily on manual content analysis (Chuan et al., 2019; Ouchchy et al., 2020). While hand coding can capture more nuances in languages compared to machines, this approach is subject to researcher bias and requires large amounts of manpower (Walker & Ophir, 2019). Automated dictionary methods, on the other hand, can detect a list of predefined keywords in large-scale text data within seconds using a computer program. However, this approach still relies on the premise that the predefined keywords represent the topic under investigation (Guo et al., 2016). Topic modeling has been widely used in AI research (Bunz & Braghieri, 2022; Curran et al., 2020; Vergeer, 2020). The method explores the latent semantic structure of text documents without predefined inputs by researchers (Blei et al., 2003; DiMaggio et al., 2013). It has been found to be a powerful technique for exploring the overarching categories and frames in discussions surrounding a communication phenomenon (Maier et al., 2018; Ylä-Anttila et al., 2022; Puschmann & Scheffler, 2016). Because it is unclear whether topic modeling outputs can be seen as mere frames, prior framing research has used semantic network analysis and other clustering methods to group sub-groups identified by topic models into meta-categories to represent higher-level frames (Ophir et al., 2021; Hase et al., 2020; Matthes & Kohring, 2008).