IEEE Technology and Society Magazine - March 2022 - 77

Figure 4. Is a healthcare worker's life
more valuable? Well, this is the option
that respondents usually prefer...
Figure 3. Illustration of a dilemma of the
autonomous cars: who to run over?
Teaching machines to make hard
choices
The domain of autonomous vehicles is very present
in the media, so it is very appropriate to make the
Moral Machine project well known throughout the
world. But the same type of techniques and way of
reasoning developed here to solve moral dilemmas
could be applied not only to autonomous vehicles,
but to many other different problems: who to select
for a job, who to grant parole to, who to choose as
a recipient of a transplanted organ, and so on (of
course, one may ask whether ethics consists primarily
in the resolution of dilemmas, as if it were the resolution
of problems of geometry... but let us leave
that now). To address the driving dilemmas they
face, the experts try to learn from the answers that
ordinary people would give to these questions.
This is the idea. Since we do not know how to
explicitly program the code of ethics of an autonomous
vehicle, let us ask people: what would you
do? In this situation. And in this other one. And so
on. Information is extracted from the set of answers,
until a set of patterns of behavior is constructed. The
Moral Machine project has dealt with an impressive
number of respondents: they gathered 40 million
decisions in ten languages from millions of people
in 233 countries and territories [9].
Something similar is done in other domains: as we
do not know how to program a set of explicit rules
to recognize a human face, we ask people to recognize
faces; we also ask them to distinguish if this
one is smiling, if this one is angry, sad, or worried.
Even body gestures can be recognized in this way: a
friendly gesture between friends who have not seen
each other for a long time or between two people
March 2022
who close a deal, a threatening gesture... These are
all techniques that are already well known and that
are working: let us try, then, to apply them to extract
moral knowledge from people's responses.
We see a concrete example in Figure 3: who
should be hit by the autonomous vehicle, the woman
or the man? And so on with various situations, some
more complex than others, but always in the form of
a dilemma A versus B: men versus women, old people
versus children, cats versus dogs, children versus
cats, passengers versus pedestrians, and so on.
Among the little figures that we can select if we
want to design our own scenario, there is even a
healthcare worker (man or woman) that is easily
identified by her case with a white cross (see
Figure 4). Maybe she carries the coronavirus vaccine
and by saving her we could save hundreds of thousands
of people. Should this influence our decision?
Problems of machine learning
Explainability: Machine learning explores the
construction of algorithms that can learn from data,
by building a model from example inputs in order
to make data-driven predictions or decisions, rather
than following strictly static program instructions
[10]. What are the problems with this way of addressing
ethical questions? First, there is the problem of
explainability, which is of great concern to researchers.
In machine learning (which is only one branch
of AI), the end result is a formula, an algorithm, for
recognizing the face, the gesture, the writing. But it
is not possible to explain why it works. There is no
other justification for the formula than it is having a
very high success rate, its effectiveness (not at all a
surprise, since learning algorithms are designed and
tested specifically to achieve a high success rate).
And of course, when the decision has a heavy ethical
77
IEEE Technology and Society Magazine - March 2022

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