IEEE Technology and Society Magazine - June 2021 - 78

Feature
contribution is presented in [29], stating that the
ethical questions related to the usage of AI can be
organized into three interconnected levels. The first
level involves laws to govern Al-related activities,
including public standards backed up by public
institutions and enforcement mechanisms, which
claim to be morally binding on all citizens in virtue
of their formal enactment. Some efforts discussed in
the " Legal perspective " section can be seen as examples
of this. However, this evades the problem that
not all of the socially entrenched standards that govern
our lives are legal standards. We rely not only on
the law to discourage people from wrongful behavior,
but also on moral standards that are instilled in
us from childhood and reinforced by society.
The second level is the social morality around AI.
The definition of such a morality is problematic as
it involves a potential infinity of reference points, as
well as the cultivation of emotional responses such
as guilt, indignation, and empathy-both of which
are effects of human consciousness and cognition
[29]. The third and final level includes individuals
and their engagement with AI. Individuals and associations
will still need to exercise their own moral
judgment by, for instance, devising their own codes
of practice. However, how these levels can be operationalized
(or to what extent) from a technical AI
point of view is not yet clear.
Open challenges
Addressing and attesting digital discrimination
and remedying its corresponding deficiencies will
remain a problem for technical, legal, social, and
ethical reasons. Technically, there are a number
of practical limits to what can be accomplished,
particularly regarding the ability to automatically
determine the relationship between biases and discrimination
and how to translate the social realities
into machine code. Current legislation is poorly
equipped to address the classificatory complexities
Table 1. Summary of challenges for the different perspectives.
arising from algorithmic discrimination. Social inequalities
and differing attitudes toward computation
further obfuscate the distinction between bias and
discrimination. From an ethical perspective, existing
moral standards need to be reassessed and frequently
updated in light of the risks and benefits AI
might pose.
In sum, the design and evaluation of AI systems
are rooted in different perspectives, concerns, and
goals (see Table 1). To posit the existence of a predefined
path through these perspectives would be
misleading. What is needed, instead, is a sensitivity
to the distinctions concerning what is desirable
AI implementation, and to a dialogical orientation
toward design processes. Finding solutions to discrimination
in AI requires robust cross-disciplinary
collaborations. We conclude here by summarizing
what we believe to be some of the most important
cross-disciplinary challenges to advance research
and solutions for attesting and avoiding discrimination
in AI.
How much bias is too much?
Whether a biased decision can be considered
discriminatory or not depends on many factors,
such as the context in which AI is going to be
deployed, the groups compared in the decision,
and other factors like a tradeoff between individualist-
meritocratic and outcome-egalitarian values.
To simplify these problems, technical implementations
tend to borrow definitions from the legal literature,
such as the thresholds that constitute prima
facie evidence of discrimination, and use it as a
general rule to attest algorithmic discrimination.
Yet this cannot be addressed by simply encoding
the legal, social and ethical context, which in and
of itself is nontrivial. Bias and discrimination have
a different ontological status: while the former may
seem easy to define in terms of programmatic solutions,
the latter involves a host of social and ethical
issues that are challenging to resolve from a positivist
framework.
Critical AI literacy
Another challenge is the need for an improvement
in critical AI literacy. We have noted the
need to take into account the end-user of AI decision
making systems, and the extent to which
their literacy of these systems can be targeted and
improved. In part, this entails end-user knowledge
78
IEEE Technology and Society Magazine
IEEE Technology and Society Magazine - June 2021

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