IEEE Technology and Society Magazine - March 2022 - 78

Feature
Figure 5. Lesson from AI: good and evil is
not what the majority says.
load, the fact that one cannot provide reasons for it
is a serious, very serious problem.
The lack of explainability is especially severe in
the field of neural networks, but it is also present in
other AI techniques, such as rule induction (also
known as rule inference). Rule induction is a kind of
learning technique that process input training data
and produce a set of IF-THEN rules to model the
desired behavior. In this case, the training data would
be the different situations or dilemmas presented to
the respondents, together with their answers (who to
run over). It has been argued that rule induction has
the advantage over neural networks that the decision
algorithm is presented as a decision tree instead
of a black box. This human-readable form would
improve interpretability and explainability. But in
fact, as long as the decision tree becomes as complicated
as necessary to cover the maximum number of
instances in the training set, it loses all its purported
" reasonableness " and " simplicity. " In the end, it is
another form of mathematical formula.
Bias: Second, another problem that arises over
and over again is the problem of bias. Ethical algorithms
have to avoid bias, whatever it is. Biases
against women, against African Americans, against
those who dress unconventionally, and so on. But
then, what if the population we are surveying in
the experiment is biased? We will be reproducing
the majority bias and that is not acceptable: if the
majority is biased, it is not enough to imitate it. The
bias must be avoided, no matter what the majority
says. In other words, the search for equity in social
relationships demands that biased judgments
are avoided.
This highlights a very interesting point: we know
(believe?) that bias, or being biased, is a bad thing,
regardless of what the majority says. In other words,
good and evil is not what the majority says, it is
78
beyond that (see Figure 5). Moreover, even if certain
biases may be beneficial in promoting equity,
then the distinction between beneficial and detrimental
biases requires a referent beyond the biases
themselves. This is not an original consideration: it
is at the very origins of ethical reflection in Greek
philosophy (e.g., Plato's warning against the risk of
making choices based on false opinion rather than
knowledge, or Aristotle's careful distinction between
the common interest of society and the interest of a
majority [11]). What is interesting is that AI brings
this consideration back into focus.
But then, if it is not what the majority says, how
do we know what is right and what is wrong? A
very difficult problem... We are not going to claim
that we have the answer, ready to be explained in
three lines. However, we do dare to say that giving
up trying to know good and evil in themselves-and
this " in themselves " means that they are beyond the
opinion of the majority-is a serious limitation for
our research, our educational programs, our social
and public policies, and so on.
Sample selection: And third, closely related to the
previous one is the problem of sample selection. The
selection of people we ask: who would you run over?
That is, who are we asking, ordinary people or sadists?
And why should not sadists be included in the survey?
Are not we biasing the sample? And how do we know
who are ordinary people, and who are sadists?
If avoiding sadists is a concern, this means that
we apply the ethical criteria already in the selection
of the sample: it is a priori, it is previous to the
experiment. Although we do not know it perfectly,
we already know, in a way, what is good and what
is bad before doing the experiment. That is why we
exclude sadists from the sample.
Ethics, majorities, and critical thinking
In short, our claim is that ethics does not consist
in imitating typical or majority behavior.
We humans do not teach ethics like that, nor do
we want it to be taught like that. If a national or
regional government were to include in its educational
programs an ethical approach such as the
following: " children must imitate the majority " ...
would we not rebel with tremendous indignation?
The very seed of ethics is critical thinking, non-conformity
with dominant thinking, the commitment of
one's conscience to recognize for oneself what is
right and what is wrong.
IEEE Technology and Society Magazine
IEEE Technology and Society Magazine - March 2022

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - March 2022
