Ethics of artificial intelligence in practice. Issues and limits

D. thesis defense of Jean-Marie John-Mathews :
Ethics of artificial intelligence in practice. Issues and limits

REGISTRATION 
Registration is mandatory 

Presentation of the subject :

While Artificial Intelligence (AI) is increasingly criticised because of the ethical issues it raises, a set of tools and methods have emerged in recent years to standardise it, such as bias mitigation algorithms, fairness metrics, explanation generation, etc.). These methods, known as responsible AI, must adapt to algorithms that are fed with increasingly granular, voluminous and behavioural data.

In this research paper, we describe how AI claims to compute the “world” without the use of the normative categories we usually use to formulate critiques. How can we normalise an AI that claims to be below the norms?

To answer this question, we have developed, from the technical literature in the discipline of AI ethics, a device for norming models in order to deal with issues of discrimination, opacity and privacy. Secondly, we have formulated four empirical and theoretical critiques to highlight the limitations of technical tools for ethics.

First, we show the limits of methods for generating a posteriori explanations of decisions made by so-called ‘black box’ AIs. Second, we show the difficulty of norming AI with so-called empirical explanation methods, assuming the ability of AIs to reveal their own biases. We then show, drawing on Boltanski’s pragmatic sociology, that the methods for combating discrimination tend towards a system of expert domination in which AI constantly modifies the contours of reality without offering any outlet for criticism. Finally, we show that AI is more generally part of an extensional movement diluting the role of institutions that have the power to stabilise social criticism while providing a grip on the world.

These four empirical and theoretical critiques finally allow us to adjust our first proposal for norming AI. Starting from a technical tool, we finally propose an open and material enquiry that allows us to constantly update the question of ends and means within AI collectives.