Ethical and safety considerations in automated fake news detection

This paper highlights ethical issues in automated fake news detection and calls for caution when deploying tools to automatically detect mis/disinformation in real-life settings. We argue that the potential harm to information consumers caused by an automated tool making a mistake requires us to better understand the mistakes that can be made. We implement three proposed detection models from the literature that were trained on over 381,000 news articles published over six months. We test each of these models using a test dataset constructed from over 140,000 news articles published a month after each model's training data. Articles in the test dataset could come from any outlet, no matter if that outlet was labelled during training or never used during training. We used these data to explore and understand two specific problems with algorithmic fake news detection, namely Bias and Generalisability. These problems arise from the models' training, design, and the inherent unpredictability of news content. Based on our analysis, we discuss the importance of understanding how ground truth is determined, how operationalisation may perpetuate bias, and how the simplification of models may impact the validity of predictions. We offer avenues for future research.