Navigating the Impossibility of Universally Ethical AI -- a Practical Guide for Researchers

Keywords: AI ethics, fairness

TL;DR: A survey of different approaches to AI ethics and why they contradict, followed by discussion on how democratic approaches and transparency can help overcome some of these challenges.

Abstract:

The area of ethics is full of trade-offs, but this is sometimes ignored in Machine Learning research. I argue that there can not be a single way to create ethical AI, agreed upon by humanity at large, but that this can be handled by considering a broader range of stakeholders and clearly stating assumptions, values, and trade-offs. I begin this paper with a survey of different current approaches to AI ethics and some of their drawbacks. I go on to highlight the impossibility of a universally ethical AI based on the inherent contradictions in plural human values and directly contradicting definitions, as well as to provide some discussion on the authoritarian connotations of aims to find a singular true morality.

Then, some approaches to handling this impossibility through a democratisation of the problem and clear communication are proposed. Lastly, a practical guide to navigating moral issues in machine learning research -- based on dialogue, transparency, and conscious trade-offs -- is proposed in the form of four basic principles and a checklist.

Serve As Reviewer: ~Lovisa_Eriksson1

Submission Number: 24