Implementing ethical principles in AI: an initial discussion

Abstract In recent years, there has been a lot of discussion around ethics in IT and AI. Many researchers and organisations have proposed guidelines to address privacy, fairness, and explainability challenges for creating trustworthy AI. In this paper, we discuss ethical principles in the context of AI and their significance in developing trustworthy AI solutions. We consider the problem of the categorisation of ethical principles in IT. We concentrate our discussion on privacy, fairness, and explainability. These principles, we believe, meaningfully contribute to the trust of AI systems. We overview the available privacy regulations in the EU and US. We also look at how to achieve compliance with them, including private data detection, data anonymisation techniques and toolkits. From a practical perspective, we analyse fairness and bias problems. We discuss the issue of fairness assessment and metrics. To improve the fairness of AI solutions, an enormous number of techniques have been developed. We also focus on fairness improvement techniques and a few popular toolkits in which these techniques are implemented. Explainability is another ethical principle discussed. It is one of many socially important properties, as it ensures understanding of AI system decision-making and transparency in inspection. Ensuring explainability is important for high-risk applications in healthcare, finance and criminal justice. Finally, we outline approaches that help in the level of explainability. With this review and analysis, we contribute to the knowledge of available techniques and toolkits that can be used by AI practitioners as an initial step in implementing ethical principles into AI solutions.