Abstract
Typically tool use poses few confusions about who we
understand to be the moral agent for a given act. But when
the "tool" becomes a computer, do people attribute moral
agency and responsibility to the technology ("it's the
computer's fault")? Twenty-nine male undergraduate
computer science majors were interviewed. Results showed
that most students (83%) attributed aspects of agency --
either decision-making and/or intentions -- to computers. In
addition, some students (21%) consistently held computers
morally responsible for error. Discussion includes
implications for computer system design.
Keywords:
Computer agents, computer ethics,
intelligent agents, social computing, social impact.
Introduction
Medical expert systems. Automated pilots. Loan approval
software. Computer-guided missiles. Increasingly,
computers participate in decisions that affect human lives.
In cases of computer failure, there is a common response to
"blame the computer." Is this a sincere instance of
attributing moral agency to a computer, or a superficial
verbal response that simply appropriates moral language?
To investigate this question, this study examined computer
literate individuals' reasoning about computers as moral
agents.
METHODS
Twenty-nine male undergraduate computer science majors
from a leading research university in California (mean age
= 23:1) participated in a one and a half hour interview about
their views on computer agency and moral responsibility
for computer error.
Note:
A considerable effort was made to interview equal
numbers of females and males; however, a low enrollment
of female computer science majors made this goal
unfeasible.
The interview contained questions in three general areas:
(1) Students' views of computer agency (the capability to
make decisions and the capability to have intentions). (2)
Students' assessments of computer system characteristics
and limitations. And (3) students' judgments of moral
responsibility for two scenarios that involved delegation of
decision-making to a complex computer system. One
scenario involved a computer system that administers
medical radiation treatment, and due to a computer error
over-radiates a cancer patient. The second scenario
involved a computer system that evaluates the
employability of job seekers, and due to a computer error
rejects a qualified worker. For each scenario, three
conditions were investigated: a fully automated computer
system that entails no human intervention; a token human
intervention in which a person with little authority and
status in the organizational hierarchy and little content area
expertise operates the computer system (e.g., a hospital
orderly in the radiation treatment scenario); and a non-token
human intervention in which a person with authority and
status in the organizational hierarchy and content area
expertise oversees the use of the computer system (e.g., the
attending physician in the radiation treatment scenario).
A coding manual was developed from half of the interviews
and then applied to the remaining half of the data. To
insure reliability of the coding scheme, an independent
scorer trained in the coding manual recoded 28% of the
data. Intercoder reliability for evaluations was 96%, for
content responses 97%, and for justifications 74%.
Non-parametric statistics were used to analyze the
categorical data. The McNemar statistic was used to
determine a change in students' evaluations across measures
(e.g., evaluation of blame across conditions). The amount
of blame students' assigned to each potential agent was
treated as score data. Then matched-pair t-tests were used
to determine differences in students' assignments of blame
across agents and conditions.
RESULTS
Due to limited space, only a few of the results will be
presented here.
Computers as Agents
The capability to make decisions and the capability to have
intentions were used to assess students' views on computers
as agents. Seventy-nine percent of the students judged
computers to have decision-making capabilities and 45%
judged computers to have intentions. Eighty-three percent
of the students attributed at least one of the two capabilities
to computers; 41% attributed both capabilities.
Furthermore, when students attributed only one aspect of
agency to computers, they were more likely to attribute
decision-making than intentions (p<.006).
Students' reasons for their assessments were also obtained.
In justifying their positive or negative assessment of
computer decision-making, virtually all students (95%)
appealed to computers as deterministic systems that make
use of rule-based or algorithmic processes, or lack free will.
For example, in support of computer decision-making one
student said, "[the computer is] deciding based on a clear
strict algorithm...it's a decision but not an open-ended one."
In contrast but also drawing on the idea of computers as
deterministic systems, to buttress a negative assessment
another student said, "the decisions that the computer makes
are decisions that somebody else has made before and
programmed into the computer....it can analyze its input and
take various actions depending on what the nature of the
input is, but somebody has already told it how to proceed in
the case of various inputs." Thus, students shared a view of
computers as deterministic systems, but differed in their
assessments as to whether or not deterministic activity
constitutes genuine decision-making.
Students drew on a largely different set of reasons to
support their assessments of computer intentions. Of the
students who judged computers to lack intentions, 36%
appealed to deterministic systems, 14% to emotions, 7% to
consciousness, 7% to the soul, and 36% provided
unelaborated responses. In many of these cases students
referred to the absence of qualities in computer systems
such as a lack of consciousness (e.g., "The program is not
actually knowing... it's like a level of consciousness...it's
just a computer that executes these lines of code...so there's
no intention on the part of the program."). In contrast,
students who judged computers to have intentions
encountered difficulty explicating their reasons. Although
probed to the same degree as students who did not attribute
intentions to computers, all of these students (100%)
provided vague, unelaborated justifications that often did
little more than reassert their assessment.
While the above findings overall provide a positive
portrayal of computers as agents, students also judged
computers to be different than humans along similar
dimensions. Of the students who judged computers to have
decision-making capabilities, 100% judged computer
decision-making to be different from human decision-
making. Similarly, of the students who judged computers
to have intentions, 77% judged computer intentions to be
different from human intentions (Z=2.30, p<.05).
Responsibility for Computer-Error
Overall, students perceived the two scenarios -- on radiation
treatment and employment rating -- as similar. No
significant differences were found between the scenarios for
corresponding agents and conditions in students'
evaluations of who or what to blame.
Roughly one-fifth of the students (on average 21%)
consistently blamed the computer system itself for the
computer-based error. No significant differences were
found across the three conditions and two scenarios for
students' evaluations of blame and the amount of blame.
However, the amount of blame finding should be
understood with caution as only those students (n<=6) who
blamed the computer were assessed for the amount of
blame.
A central concern of this study is how students understand
computers to be accountable, if at all, for computer error.
Thus, it is useful to examine students' reasons for blaming
or not blaming computers in relation to their reasons for
blaming or not blaming people (the computer system
designer, the computer system's human operator, and the
organization's administrators). Averaging across conditions
and scenarios, virtually all of students' justifications for
blaming the computer (96%) referred to the computer's
participation in the sequence of events that led to harm. In
contrast, the large majority of students' justifications for
blaming people (80%) referred to failing to meet some
commonly expected reasonable level of performance (e.g.,
negligence). When students did not assign blame,
differences were also found among the justifications
students used for computers and for people. Again,
averaging across conditions and scenarios, virtually all of
students' justifications for not blaming the computer (97%)
referred to qualities of computers that diminish its agency
and thus undermine computers as being the sort of thing
that can be blamed. Notably, the appeal to diminished
agency was used exclusively in reference to computers. In
contrast, students' justifications for not blaming people
primarily referred to adequately meeting commonly
expected levels of performance (55%) and deferring to an
authority perhaps due to habit, lack of autonomy, or the
authority's greater power or knowledge (41%).
DISCUSSION
The data reported above joins a growing body of research
[1, 2,] that suggests people, even computer literate
individuals, may at times attribute social attributes to and at
times engage in social interaction with computer
technology. Some researchers argue that as good designers
we ought to exploit this psychological phenomena to build
systems that actively engage users in a social relationship
with the technology. Much of the work on computer agents
and intelligent agents is of this vein. The results reported
here, however, should give us pause. For the results
suggest that even some computer literate individuals hold
computer technology at least partly responsible for
computer error. If this finding is correct, a different design
strategy is in order. It would follow, for example, that
designers should communicate through the system that a
(human) who -- and not a (computer) what -- is responsible
for the consequences of the computer use.
Acknowledgments
We thank Sara Brose and Sue Nackoney for help with the
coding, reliability, and analysis of the data. This research
was funded in part by the Clare Boothe Luce Foundation
and by a Natural Sciences Division Research Grant from
Colby College.
References
1. Nass, C., Steuer, J. and Tauber, E. R. Computers Are
Social Actors, in Proc. CHI '94 Human Factors in
Computing Systems (Boston, April 24-28, 1994), ACM
Press, 72-78.
2. Walker, J. H., Sproull, L. and Subramani, R. Using a
Human Face in an Interface, in Proc. CHI '94 Human
Factors in Computing Systems (Boston, April 24-28, 1994),
ACM Press, 85-91.