IEEE Technology and Society Magazine - June 2021 - 67

What is " ethics " in AI ethics?
Even though the term " ethics " cannot have a
straightforward definition, nor can it entertain anyone
who wishes to define, philosophers differentiate
between " morality " and " ethics. " Morality is descriptive,
in the sense that covers the actual behavior of
individuals, while ethics is prescriptive in nature and
dictates what individuals should do. In other words,
morality answers the question: what do people do?
and ethics answers the question: what should people
do? Many AI researchers and practitioners have never
heard of this distinction, nor even take any " ethics 101 "
class or lesson, even though they build AI-powered
systems or socio-technological systems that interact
in the world and can have an impact on individuals
and society as a whole. What is understood by ethics
in the computer sciences curricula is not the same as
what is understood by ethics in philosophy, political
philosophy, and other social and human sciences. To
put it bluntly, ethics is an area of study with a long tradition
(spans several millennia) that deals with complex
concepts and actions that draw on the insights
of multiple disciplines that have emerged throughout
history such as anthropology, economics, psychology,
etc., to name a few. In other words, ethics is hard. Let
us quote in length a recent treatment in book format
about how to decide among different ethical theories
in the face of uncertainty to grasp how hard ethics is:
" As with other areas of philosophy, working out the
correct moral view often involves being sensitive to
subtle distinctions, being able to hold in mind many
different arguments for different views, and paying
attention to intuitions across many different thought
experiments. It also involves difficult questions about
how to weigh different theoretical virtues, such as
simplicity and elegance against intuitive plausibility.
Correctly balancing all these different considerations
is extremely difficult, so even when we come to a
firm stance about some ethical view, we should not
always expect that our reasoning is error-free " [4, p.
11]. " Ethics " in AI ethics for the promotion of SDGs or
any other social good aims to investigate and identify
critical ethical issues when building AI capabilities,
and the main concern of ethics is to identify product
vulnerabilities and recommend strategies for prevention
based on sound ethical analysis. Laypeople and
people outside the professional field of philosophy
usually use the terms ethics, morality, and even law
interchangeably. As we have shown above, ethics
and morality are two distinct concepts. To be more
June 2021
specific, morality is a set of norms, values, principles,
etc., to behave correctly which are shared by a group.
On the other hand, ethics is the philosophical study
of norms, values, principles, etc., and what it means
to behave correctly or wrongfully. And law is a set of
formal rules which confers duties and rights to organize
society. What is " ethics " in AI ethics is a question
of value alignment as well [5]. The main problem
with the alignment of values or ethically aligned AI
is how we decide which values to implement given
that there is no agreed definition of what ethics is and
of course we live in pluralistic societies where people
differ in which values are most important. But there
is also a problem with what we have to understand
by AI. AI can be understood as the science and engineering
that seeks to create intelligent machines. For
the purpose of this article, we understand by AI any
artificial system that seeks to fulfill its objectives, whatever
these may be. AI value alignment refers to how
we make an artificial system compliant with human
values [6]. An important aspect to consider in the AI
value alignment issue is the type of method we use
to build an AI system and the encoding of values
in those systems: the technical and the axiological.
Within machine learning, a set of techniques and
methods that allow machines to learn from data to
make predictions and improve decision-making to
create intelligent machines, there are different subfields
and each of these different subfields constrains
the type of values or ethical principles to be codified
in the design of an AI model.
* Supervised learning (SL): Train an AI model to
perform a task with labeled data.
* Unsupervised learning (UL): Train an AI model to
perform a task with unlabeled data.
* Reinforcement learning (RL): An agent learns to
maximize a reward signal from the environment.
Each of these technical methods for building an
AI seems to correspond to an established classical
ethical doctrine [5]. For example, SL is Kantian or
deontological in its characteristics. SL follows the
Kantian maxim: " Act only on that maxim through
which you can at the same time will that it should
become a universal law [of nature] " but in this case
applied to an AI system. In other words, the programmer
or the engineer labels the data to train the AI
model and therefore codes some values or principles
considered by him as reasonable and universal
that have to guide the system's behavior. Meanwhile,
67
IEEE Technology and Society Magazine - June 2021

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