Who is to blame for AV crashes? Public perceptions of blame attribution using text mining based on social media

Abstract

Autonomous vehicles (AVs) are transitioning from the laboratory to the market phase. However, frequent AV crashes hindered the commercialization. Since the human driver and AV systems jointly control the vehicle, responsibility is difficult to attribute. Previous research primarily investigated the blame attribution in AV crashes from an ethical or legal perspective using traditional questionnaires based on vignette-based scenarios, but reached contradictory conclusions. Public perception, as a significant source of law, is often overlooked. The research extracted 90,426 valid comments on Chinese social media platforms (Sina Weibo and TikTok) about AV crashes from April 2021 to January 2025, breaking down the limitations of traditional attribution research with novel data sources. Bidirectional Encoder Representations from Transformers (BERT) model exceeds other deep learning models in classification accuracy of blame attribution comments. The Latent Dirichlet Allocation (LDA) results revealed an interesting topic Publicity, expanding the targets of traditional research. Regarding the topic AV system, the public appeared to have difficulty in classifying AV levels, with the highest blame volume and most negative sentiments. Importantly, we found certain connections among topics that chaotic publicity results in biased public perceptions of AV systems, thereby leading to over-trust in AV technology, which ultimately increases crash rates.