COUNTERFACTUALS AND ARTIFICIAL INTELLIGENCE EXPLICABILITY PROBLEM

Abstract

In the first part of the text, I present the most important ethical problems in the application of AI, such as perpetuating existing prejudices and discrimination and increasing injustice and inequality, explainability, privacy, transparency, responsibility, and autonomy. The problem of explainability refers to the fact that people do not sufficiently understand why AI makes certain decisions because the principles and processes behind those decisions are not sufficiently clear to them. Recently, several authors suggested that counterfactual conditionals could be used to increase explainability. The second part is a review of the most important theories of counterfactuals.
The third part is a review of the application of counterfactuals in explanations. Understanding the reasons for making particular decisions and why that decision might change under different conditions is the key to building trust in AI models. Explanations can serve many purposes: to inform and help the user understand why a particular decision was made, to provide grounds to contest adverse decisions, and to understand what can be changed to achieve a desired result in the future. In the current literature, “explanation” includes opening the “black box” to provide insight into the internal decision-making process of algorithms. However, explaining the functionality of complex algorithmic decision-making models and their rationale in specific cases is a technically challenging problem. In contrast, counterfactual explanations describe dependency on the external facts that led to that decision. Thus, they can, in principle, be offered without opening the “black box”, what is seen as their main advantage.