Research Exploration of Artificial Intelligence: The Black Box

Abstract

The fact that the explosion of Artificial Intelligence (AI) has given way to the popularity of complex models, particularly deep learning models, which are more often than not opaque black boxes. This paper is a critical discussion of issues arising with these non-transparent systems, especially on ethics in all its applications, trust and regulatory concerns about high-stakes systems like healthcare, finance and criminal justice. We shall provide a critical assessment of the most prominent explainable AI (XAI) techniques such as LIME, SHAP, and counterfactual reasoning and compare the extent to which they succeeded in making complex models interpretable. Also, we evaluate contextual risks linked to different application areas and stress the difference between interpretability requirements in different industry sectors. To conclude, the paper offers constructive guidelines to future AI research, as well as a call to support interpretable-by-default models, comparable interpretability measures, and a more in-depth consideration of ethics and law. Based on this, the paper would contribute to good AI practices that are based on accuracy, transparency, and building trust in multiple fields of different disciplines.

Ähnliche Arbeiten