Lupine Publishers| Journal of Robotics and Mechanical Engineering
Introduction
The issue of Facebook
moderators made topical headlines in 2019. When confronted with horrifying
images such as terrorist assaults filmed online, moderators are required to be
extremely reactive in order to eliminate the scenes violating human dignity from
the social network as rapidly as possible. In [1], my colleagues and I explored
whether it was possible or not to model human values such as virtue. Modeling
could eventually make it possible to automatize all or part of the moderators’
arduous and ungrateful work. In a first part, this article deals with the need
for a reflection on ethical Artificial Intelligence (AI). After providing a
definition of what AI is, I then discuss how ethical rules could be implemented
in a system using AI. In a second part, I ask myself whether it is possible to
embed moral and ethical principles. Using a utility function can help agents to
make ethical decisions, but it is also important to be cautious about the
limitations of such a sometimes-simplistic approach. Ethics must be appreciated
in the social and technical context in which it is first developed and then
implemented [2].
The Need for Ethical Artificial Intelligence
Artificial intelligence
A definition AI can be
described by a universal triad, namely the data brought by the environment, the
operations defined as the logic which mimic human behavior and finally, a
control phase aiming at retroacting over its previous actions. Its definition
is essentially based on two complementary views: one focused on the behavior
and how it acts, especially as a human and the other which emphasizes the
reasoning processes and how it reproduces human skills [3]. However, both
points of view insist on the rational behavior that an AI must have. Moreover,
it is important to pay attention to which kind of AI we are dealing with:
strong or weak AI [4]. Weak AI, also known as narrow AI is shaped by behaviors
answering to observable and specific tasks that may be represented by a
decisional tree. On the other side, the strong AI, or artificial general
intelligence, can copy human-like mental states. For this type of AI, this
means that decision abilities or ethical behavior are issues that need to be
taken care of. Finally a strong AI could find the closest solution to the given
objective and learn with an external reversions. The latter is the one that is
posing unprecedented problems that researchers are just starting to study. In
fact, a system embedding strong AI is able to learn without human assistance or
injection of additional data since the AI algorithm generates its own
knowledge. Therefore, the exterior observer or the user of such agents will no
longer know what the AI knows, what it is capable of doing nor the decisions it
is going to take. Hence the need to establish an ethical framework that defines
an area of action and prevents the system from taking decisions contrary to the
ethics.
How to implement ethics rules ?
In order to implement
ethical rules, there are two approaches. The first one named top down approach
is based on ethical ruleabiding machines [5-6]. The strategy is to respect
unconditionally the ethical principles related to morality such as “Do not
kill”. However, without understanding the potential consequences of the
empirical decisions taken, an AI system creates numerous approximations that
are a significant drawback. This can make rules conflict even for the 3 laws of
robotics [7]; it may also lead to unintended consequences due to added rules
[8]. It should be noted that even inaction can be taken into account for
injuring humans. Moreover, the complexity of interaction between humans’
priorities may lead to interpersonal inappropriate comparisons of various added
laws [9]. The second method called bottom up, focuses on case studies in order
to learn general concepts. The case studies make it possible for a strong AI to
autonomously learn wrong and biased principles and even generalize them by
applying them in new situations it encounters. These types of approaches are
considered to be dangerous. In fact, in this process of learning, basic ethical
concepts are acquired through a comprehensive assessment of the environment and
its compliance with previous knowledge. This is done without any bottom-up
procedure. The result of this learning will be taken into account for future
decision making [10,11]. Eventually, in the case where an AI algorithm is
facing a new situation it has not encountered before, the extrapolation without
a control phase may result in perilous situations for humans [12].
Artificial Intelligence Embedding Moral and Ethical Principles
A utilitarian approach
of ethics consists in choosing in the case of a set of possibilities the
solution that leads to the action consequently maximizing intrinsic good or net
pleasure [13]. This involves quantifying Good or Evilfrom a given situation.
However, certain situations supported by ethical reasons with an empirical
study may prohibit the combined execution of certain actions. These complex
cases are at the origin of dilemmas and the ethical principles do not make it
possible to establish a preference. Therefore, autonomous agents need to be
endowed with the ability to distinguish the most desirable option in the light
of the ethical principles involved.
To achieve this goal, this article proposes in the following subsections a
method called a utility function as a mean of avoiding ethical dilemmas.
Using a utility function to help agents make
ethical decisions
In order to achieve this
goal, a number of solutions have been proposed [14]. One of them it the utility
function, also known as objective function. This function is used to assign
values to outcomes or decisions. The optimal solution is the one that maximizes
the utility function.This approach based on quantitative ethics determines
which action maximizes benefit and minimizes harm. Its objective is to make it
possible for an AI algorithm to take the right decisions particularly when it
encounters an ethical dilemma. From a mathematical point of view, the utility
function takes a state or a situation as an input parameter and gives as a
result an output which is a number [15]. This number is an indication of how
good the given state or situation is for the agent. The agent should then make
the decision that leads to the state that maximizes the utility function. For
instance, let’s take the case of an autonomous vehicle, and let’s assume that
the car is in a situation where harm is unavoidable, and that it would
inevitably hit either two men on the road or crash into a wall killing the
passenger it is carrying [16]. Based on our previous utilitarian ethics
definition, the decision that will minimize harm is the one that will lead to
kill as few people as possible. Therefore, the car should crash and kill the
passenger to save the two pedestrians because the utility function of this
outcome is the highest. The same reasoning applies for military drones when
they have to choose between multiple outcomes that involve moral and ethical
principles.
Autonomous cars
embedding AI algorithms using the utility function are not yet marketed. Some
models available to the general public have an autopilot mode that still
requires the presence of a human being behind the steering wheel who will make
a decision in case of a problem. Fully autonomous cars still ride in test
environments [17]. In the near future, the people who are likely to buy this
type of car will primarily be public institutions such as municipalities. For
instance, the city of Helsinki is testing an autonomous bus line, RoboBusLine
which carries passengers on a defined road with a limited speed and an
autonomous shuttle is also in service in Las Vegas [18]. However, these are
still prototypes in a test phase with an operator on board. The other customers
that may be interested in using autonomous vehicles are companies that make
deliveries given the advantage of automating the tasks resulting in cost
reduction and efficiency. In fact, Amazon, FedEx and UPS are investigating
solutions for driverless trucks. The utility function is currently under investigation
as an active solution to avoid ethical dilemmas non-modifying on-policy [19].
Autonomous robots are expanding, and the aim is not only to deal with ethical
dilemmas but also to reduce uncertainty by quantifying problems such as
exploration or unknown mapping; both can be stochastically defined (Shannon or
Rényi’s entropy) [20,21]. Describing and taking actions in a world incompletely
defined can be done with the help of estimators but utility functions describe
the perceptual state in line with the rules, and an active strategy can hence
be implemented. This is already done for robot vision for example [22].
Limits and dangers of the utilitarian approach
The approach previously
described and consisting in quantifying situations and assessing them with a
utility function through a model has its own limits as far as strong AI is
concerned. For weak AI, engineers at the design stage can implement decision
trees to establish rules. They can anticipate the behavior of the AI more
easily. On the other hand, as mentioned in section 2.1, advanced AI systems
learn directly from the environment and adapt accordingly. By doing so, an
external observer cannot always predict or anticipate the actions of such
systems [23]. This is true for the Alpha Go algorithm that is taking decisions
and implements strategies even experts in the game cannot understand although
it leads to an optimum solution. The intelligent agent behaves like a black box
whose internal functioning is unknown. This is particularly dangerous when it
comes to autonomous vehicles or UAV drones that put human life at stake. Using
only a utility function to decide whether or not a UAV could be used in an
armed conflict could be considered as a war crime. Indeed, it is essential to
test AI algorithms in different environments [23] and cover as many situations
as possible before they are registered for use. This involves confronting
algorithms with different situations and ensuring they behave properly by
taking the most ethical decisions possible. It will then be possible to
identify anomalies and correct them immediately.
Conclusion
In just a few years,
artificial intelligence has become a strategic issue in Europe, the USA and
China. For fundamental considerations of balance of powers between that of the
GAFAM on the one hand and ethics at the service of the greatest number of
people on the other hand, it will be crucial to develop an “ethics in context”.
This is more than computational ethics and utility functions developed by
economists. It will be possible to implement an embedded code of ethics using
artificial ethical agents, but only if ethical principles remain submitted to a
democratic deliberation involving all participants.My future research will
focus on the question of “human values and AI” at different levels:
universality, continentality (USA vs. China for example), nation-state and
local communities using AI.
