IEEE Technology and Society Magazine - Winter 2014 - 26

This speculation and discussion illustrates how
imaginative future visions of technological development can be misleading. Computer scientists use the
concept of autonomy to explain how new technologies
would work, and to provide a goal to strive for. Some
computer scientists use autonomy to refer to high-level
automation. Some use it to refer to the capacity that
some artificial agents have to navigate in environments
by relying on machine models of the environment.
Yet others see machine autonomy simply as responsiveness to an environment wherein sensors provide
input and artificial agents are programmed to respond
to those inputs. [See [21] for a discussion of different
conceptions of autonomy.] However valuable, these
interpretations of autonomy are distinct from conceptions of autonomy in moral philosophy or daily life,
where autonomy is intertwined with notions of free
will, responsibility, and intentionality. Failing to recognize the differences between these conceptions of
autonomy can lead to confusion about how artificial
agents actually work, and to unjustified inferences
about responsibility. It can lead to the unjustified
conclusion that entities with machine autonomy can,
themselves, be responsible, or that humans are less
responsible for the behavior of such entities.
As the various conceptions of autonomy suggest,
claims about the autonomy of artificial agents have to
be understood within the context in which they originate. For example, in developing policies and regulation for the use of autonomous cars on public roads, it
is not enough to simply assume that the autonomy of
these cars is self-explanatory. Autonomous cars may be
driverless cars or cars with some autonomous components, such as a self-parking function. The difference is
important among other things for policy making, since
policies would be different for each [23]. Similarly, the
prospect of increasingly autonomous military robots
draws attention away from the activities of operators, technicians, and other human actors involved in
the operation of these robots. The latest generation of
UAVs are said to be autonomous because they can take
off from and land on an aircraft carrier on their own.
However, this kind of autonomy is very far off from the
kind of autonomy that the media and popular science
articles predict [35]. That is, taking-off, landing, and
navigating is quite different from independently making targeting decisions in active combat.
Another problem with underspecified and contextless conceptions of autonomy is that they obfuscate
the relationship of control between human actors and
the technology. When the autonomy of a UAV refers
to its ability to take off, land, and navigate between
waypoints without human intervention under "normal conditions," the operator may still be required to
monitor the performance of the aircraft when the conditions are irregular. Such an autonomous technology
26

has a human operator, an operator that must be in a
position to exercise responsibility, e.g. she has to be
continuously aware of and able to understand the status of the aircraft. Specifying and contextualizing the
meaning of "autonomous" leads, thus, to paying attention to assumptions about the distribution of tasks and
control among human and non-human components.
Recommendation 3 is not just important for developers who typically present new technologies to
funders, users, and the public and may have an incentive to use hyperbole - it is important for all who communicate to the public. This includes policy makers
and legal experts, as well as researchers, scholars, and
journalists. Keeping claims about the autonomy of
artificial agents in context can contribute to a more
informed public and the possibility of better public
input to the future development of artificial agents.
Recommendation 4:
Responsibility issues involving artificial agents can
best be addressed by thinking in terms of responsibility practices.
Part of the problem in addressing responsibility
issues raised by complex technologies is that the conventional moral notions of responsibility are difficult
to apply to humans acting with technologies. These
notions place an emphasis on the relationship between
an individual's actions and capacities, and the outcome of those actions and capacities, ignoring the
contribution made by a technological environment. In
technological environments, the individual may not
have full control of the outcomes and may have had
only a partial understanding of where their individual
action would lead. The technological contribution
exacerbates the challenges of locating responsibility.
When something untoward happens, such as a drone
killing an innocent person, tracing back what went
wrong and identifying what or who is responsible can
be a daunting task, though that does not mean there is
no one responsible.
Because of the complexities of responsibility
ascription, it is useful to think about responsibility as
a set of practices [22]. That is, what or who is responsible for an outcome depends on the responsibility
practices in the context at issue. These practices are
the established ways in which groups and individuals
in a community understand, evaluate, and distribute
responsibility. Individuals and groups come to have
a responsibility or to be responsible for something as
a result of social norms and shared ideas about what
sort of behavior is expected in particular contexts and
what consequences will follow from living up to or
failing to live up to the expectations.
These norms are established and reinforced in a
wide variety of ways. For example, managers and
colleagues inform new engineers about what their
IEEE TECHNOLOGY AND SOCIETY MAGAZINE

WINTER 2014

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - Winter 2014