IEEE Technology and Society Magazine - June 2021 - 68

Feature
UL seems to follow an Aristotelian virtue ethics in the
sense that it allows the AI algorithm to find patterns
in the unlabeled data for a solution to the required
task. In other words, because this kind of ethical theory,
Aristotelian virtue ethics, is built on self-realization
and UL trying to uncover patterns in unlabeled
data is like as if the system self-realizes to achieve the
required task. For the case of RL, the parallelism with
another classical ethical doctrine, utilitarianism, is
clear. An agent (AI system) wants to maximize the
reward signal from the environment and according
to act utilitarianism a right action is the one that creates
the greatest happiness for the greatest number
of people. The tradeoff between these two aspects,
the axiological and technological, is not always easy
because they are intertwined. So, the type of AI technique
chose influences the principles and values
that will constrain the behavior of the system. But
here we enter into another fundamental problem
of a philosophical nature. The distinction between
facts and values [7]. Many of the techniques within
machine learning are trained with data (facts) but
these data do not have to contain correct " values. "
That is, even if an AI model imitates what a programmer
has labeled as correct or discovers patterns in
the data, these newly discovered patterns do not
have to be axiologically correct or just because they
imitate what a human being has done they have to
be correct or valuable. And this problem is exacerbated
when you decide to design an AI technology
the ultimate goal of which is to do social good such
as the promotion of SDGs.
Table 1. List of some specific metrics of AI research on fairness.
Source taken from Fairness and Machine Learning by Solon
Barocas, Moritz Hardt, and Arvind Narayanan.
Ethics by design
IEEE's ethically aligned design (https://ethicsinaction.ieee.org/)
puts a strong emphasis in human
wellbeing when building AI technologies. Montréal
declaration for a responsible development of AI
(https://www.montrealdeclaration-responsibleai.com/
the-declaration) defends certain values and ethical
principles that protect individuals and groups. As AI
technologies become more prevalent in society, it
becomes more than necessary to implement principles
of regulation and control of AI. But the number
of guidelines, manifestos, statements by industry,
governments, and civil society is growing rapidly [8].
And none is strong enough to be used widely by all
stakeholders. Most of these guidelines are analytically
relevant, accurate, but often have a simplified view of
ethics as discussed above. As we will present in the " HR
as a normative framework for AI technologies " section,
the best normative guide to establish ethical principles
of governance of AI technologies is the human rights
(HR) framework. But first, we would like to point out
some basic principles that some of these guidelines discuss
and that are necessary for an AI ethics.
Fairness
Fairness is central to human moral cognition but
at the same time is very controversial too. Its definition,
its origins, or the possible presence in other
species is elusive and a source of conflict between
a large number of disciplines. If we correctly understand
what fairness means our society can be better
organized [9]. Fairness is not the same as equality.
Equality means same distribution of resources
among all, while fairness means just distribution of
resources among all. The following example clarifies
the difference. An equal distribution of four apples
between two people is that each gets two. But a fair
distribution is one when one of the two has contributed
more, say, has planted the tree and cared
for it and picked the apples, he takes more apples
for the effort in contrast to the other who has done
nothing. In this sense, fairness is equal to justice or
just causes. In relation to the use of an AI system, in
order for it to be fair does not have to produce equal
outcomes. Many specific metrics in AI research on
fairness focus on a parity outcome (see Table 1) and
this is not fairness properly speaking. A recent article
[10] has reviewed the three main mathematical
definitions of fairness concluding that all of them
suffer from significant statistical limitations. This is
68
IEEE Technology and Society Magazine
https://ethicsinaction.ieee.org/ https://ethicsinaction.ieee.org/ https://www.montrealdeclaration-responsibleai.com/the-declaration https://www.montrealdeclaration-responsibleai.com/the-declaration
IEEE Technology and Society Magazine - June 2021

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - June 2021
