Ethical Biases in Machine Learning-based Filtering of Internet Communications

The use of automated systems based on artificial intelligence and machine learning for filtering and moderating online communications has become commonplace. While this allows for high levels of efficiency and fine-grained control of malicious behaviors, it could also produce unintended disparities in treatment of legitimate users. In this paper, we aim at identifying some possible field-related biases in the wellknown Google Perspective API machine learning-based engine for controlling Internet communications. For this purpose, we consider communications in the fields of health, trade, finance, and defense and build a data set collecting Twitter-based online communications of the World Health Organization (WHO), World Trade Organization (WTO), International Monetary Fund (IMF) and North Atlantic Treaty Organization (NATO). Collected data are then analyzed through Perspective API to assign them an alleged likelihood of being abusive for specific emotional concepts, referred to as attributes. Upon analysis, discrimination between the considered users is identified for all attributes. This result, although preliminary, apparently indicates that Perspective API creates discrimination for field-related content as a result of semantic biases in the data, thus highlighting the need for an ethically sound design of these systems, following an ethics by design approach.