By Thomas B. Sheridan & Raja Parasuraman
Automation does not mean humans are replaced; quite the opposite.Increasingly,humans
are asked to interact with automation in complex and typically large-scale systems,
including aircraft and air trafﬁc control, nuclear power, manufacturing plants, military
systems, homes, and hospitals. This is not an easy or error-free task for either the system
designer or the human operator/automation supervisor, especially as computer technol-
ogy becomes ever more sophisticated. This review outlines recent research and challenges
in the area, including taxonomies and qualitative models of human-automation interac-
tion; descriptions of automation-related accidents and studies of adaptive automation;
and social, political, and ethical issues.
The technological revolution ushered in by the computer has dramatically affected
many aspects of human activity—at work and at home, during travel, and while
engaged in leisure pursuits. Even more radical changes are anticipated in the next
decade as computers decrease in size and cost and increase in power, speed, and
These factors are responsible for much of the drive toward increased automa-
tion in the workplace and elsewhere. The economic beneﬁts that automation can
provide (or is perceived to offer) has motivated considerable research and
development on the technical capabilities of automation, which have been amply
documented in such diverse domains as aviation; manufacturing; medicine; road,
rail, and maritime transportation; robotics; home and entertainment devices;
and numerous others. Humans work with or are consumers of all these technolo-
gies. Consequently, understanding how human characteristics and limitations inﬂu-
ence the use (or misuse) of automation and using such knowledge to better the
design of automated systems have been the focus of considerable research over the
past two decades (Bainbridge, 1983; Billings, 1997; Jamieson & Vicente, 2005; Para-
suraman & Mouloua, 1996; Parasuraman & Riley, 1997; Rasmussen, 1986; Sarter,
Woods & Billings, 1997; Sheridan, 1992a, 2002; Wickens & Hollands, 2000; E. L.
Wiener & Curry, 1980).
In this chapter we discuss research on humans and automation. We do not pro-
vide a comprehensive review of the ﬁeld but describe recent and seminal work on the
topic. We begin by deﬁning automation and describe taxonomies and qualitative
models of human-automation interaction, including the supervisory control model,
function allocation, and the concept of human-centered automation. We then discuss
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 89
automation-related accidents and incidents associated with inadequate feedback
about system states,misunderstanding of automation, and overreliance.Subsequently,
we describe recent research on the role of trust and “etiquette”in human-automation
Because empirical human performance studies can be guided by and their results
better interpreted with quantitative models, we brieﬂy describe some of these models.
Research on adaptive and adaptable automation is described next, with discussion of
applications to driving and air trafﬁc control. We then take a look at the future and
discuss what the impact of new automation technologies might be. We close by dis-
cussing some social, political, and ethical issues that arise in considering the relation-
ship between humans and automation.
WHAT IS AUTOMATION?
The Oxford English Dictionary deﬁnes automation as follows: “1. automatic control of
the manufacture of a product through a number of successive stages; 2. the applica-
tion of automatic control to any branch of industry or science; 3. by extension, the
use of electronic or mechanical devices to replace human labor.”
The ﬁrst use of the term automation is traceable to a 1952 Scientiﬁc American arti-
cle. Today use of the term has grown beyond product manufacturing and is applied
to automatic control and instrumentation for chemical and power plants, aircraft and
air trafﬁc control, automobiles, ships, space vehicles and robots, heating and air con-
ditioning in buildings, business systems, medical devices, home appliances, military
systems, and stand-alone computers, to name only a few examples. Thus the second
meaning is still widely accepted, as is the third meaning when human labor means
mental as well as physical labor.
Mental labor is of primary importance for automation today, at least in the devel-
oped nations. Computers that interpret inputs, record data, make decisions, or gener-
ate displays are now regarded as automation, including the sensors that go with them,
even though in the strict sense none of these functions may be automatically con-
In the fullest contemporary sense, the term automation refers to
a. the mechanization and integration of the sensing of environmental variables (by artiﬁ-
b. data processing and decision making (by computers);
c. mechanical action (by motors or devices that apply forces on the environment), and/or
d. “information action” by communication of processed information to people.
Automation can refer to open-loop operation on the environment or closed-loop
control. It can apply to dynamic processes that change gradually over many months
or those that occur in milliseconds. Thus contemporary deﬁnitions of automation
refer to the gamut of processes, from the sensing of the environment to actions taken
on that environment (Moray, Inagaki, & Itoh, 2000; Parasuraman, Sheridan, & Wick-
90 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 90
TAXONOMIES AND QUALITATIVE MODELS
The Meaning of Human-Automation Interaction
What does it mean for human and automation to interact? Humans can be totally pas-
sive benefactors of automation: They purchase and use goods manufactured by
automation.They consume electrical power, water, and heating and vehicular fuels that
automation played a large role in providing. But being a passive user in this sense is
not really interacting with the automation per se.
What we mean by human-automation interaction is the circumstances in which
people (a) specify to the automation (necessarily a computer of some sort) the task
goals and constraints (do X but avoid doing Y) and trade-offs between the goals and
constraints; (b) control the automation to start or stop or modify the automatic task
execution; and (c) receive from the automation information, energy, physical objects,
Simple examples of item 1 are people pushing the ﬂoor button on an elevator or
setting the controls on their washing machines, and people setting the speed controls
on their automobile cruise control system. More complex examples of item 1 include
a. pilots programming their ﬂight management systems using a digital keypad or using spe-
cial command language to have the autopilot take them to a new altitude and heading,
ﬂy to a series of designated waypoints, and enter the landing pattern at a distant airport;
b. machinists similarly programming a numerically controlled machine tool to make a metal
part in a series of machining operations; and
c. space engineers programming movements of a robot arm on a Mars rover.
The conditions for the automation to start or stop or modify its program in the sec-
ond item may be, for example, clock time or when sensors in a chemical plant indi-
cate that a certain temperature has been reached, or when a robot has made contact
with an object. Operators may abort automatic execution, and/or human takeover may
be imposed when the human perceives a problem.
Much automation these days is used to give the operator information,either by warn-
ing or alarm display or by expert system to give advice. The information can be used by
the human operator to reach a decision about some aspect of the system and to take
action if necessary. But as in the case of decision aids, the automation may also provide
a recommended choice or course of action. As stated in the third part of our deﬁnition,
automation outputs can be energy (to move the whole body, as does an automobile or
aircraft, or a part of the body, as does a prosthetic arm), an object (as with a vending or
automatic teller machine), or a substance (as with an automatic drug delivery machine).
Supervisory Control Paradigm
The central role for humans in automated systems is to undertake what is called super-
visory control. This is a new relation between the human and the machine, as an auto-
matic machine may be said to be intelligent in some rudimentary sense. The new form
Human-Automation Interaction 91
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 91
of interaction differs dramatically from the traditional interaction of the human with
tools and devices that possess no intelligence, in which all sensing and control were
done by the human operator. This new relation was ﬁrst called human meta-control
(Sheridan, 1960) and later human supervisory control (Moray, 1986; Sheridan, 1992a
Sheridan & Verplank, 1978). The human supervisor of the automatic machine was
likened to the human supervisor of intelligent but subordinate humans, whereby the
supervisor issued instructions (goals, constraints, and plans) and the subordinates exe-
cuted those instructions using their own memories, built-in programs, sensors, and
Figure 2.1 shows the relation of the human supervisor to the (typically multiple)
computer-controlled machines performing simultaneous tasks (shown below the
human ﬁgure). The several “task” boxes represent what might be different systems
within a factory or power plant or aircraft, different degrees of freedom of a single
robot, or multiple robots or unmanned vehicles the human supervises.
Above the human ﬁgure are boxes representing the ﬁve functions of the supervisor:
1. plan off-line,
2. teach the automation,
3. monitor the automation’s execution of the plan,
4. intervene to abort or assume control as necessary, and
5. learn from experience.
92 Reviews of Human Factors and Ergonomics
Figure 2.1. Generic functions of supervisory control.
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 92
To plan involves having some mental or computer model of the physical system to be
controlled, having some trade-off between performance objectives that can be satisﬁced
(made acceptable), and formulation of a strategy for doing the task. To teach means to
decide on a desired control action and to communicate commands to the automation
to implement that action. To monitor means to allocate attention among the appropri-
ate displays or other sources of information about task progress and from these to esti-
mate the current state (vector) of the system (to maintain situation awareness). To inter-
vene means that if an abnormality of sufﬁcient magnitude is detected and diagnosed,
the human will either reprogram the automation or may even take over and exercise
manual control. Learning from knowledge of the results, like planning, is an out-of-the-
loop human function and feeds back into planning the next phase of supervision.
In supervising a single or several parallel automated systems, as shown at the bot-
tom of Figure 2.1, the human brings to bear whichever of these ﬁve functions is appro-
priate. At any instant of time the most appropriate function will typically differ from
one task to another.
Sequential Stages of an Automated Large-Scale System
A large-scale system typically involves four classes of task, each a subtask or stage of
a larger process (Figure 2.2):
a. the acquisition of information,
b. the analysis of that information,
c. the decision about actions to take based on that information, and
d. the implementation of that action.
Automation at each stage involves all ﬁve of the supervisor functions described earlier
(a 4 3 5 matrix). For example, air trafﬁc control involves (a) the acquisition of radar
information on location, ﬂight plans and identity of many aircraft, weather informa-
tion, and so on. It requires (b) that appropriate information then be combined, ana-
lyzed, and displayed to the air trafﬁc controller. Then it requires (c) that decisions be
made as to speed, heading, and altitude for different aircraft to maintain safe separa-
tion and bring the aircraft safely though a sector of airspace or to land or take off.
Finally, it requires (d) a means to get the pilots (and aircraft) to cooperate and exe-
cute the instructions given.
One can think of these four stages as the computer-controlled task boxes at the
lower part of Figure 2.1. But they are not four independent tasks; they are four tasks,
and each successive task depends totally on the previous one, as implied by Figure 2.2.
These four stages can be automated, as Figure 2.1 implies. But they need not be fully
automated. Some can be automated to a greater degree and others to a lesser degree
(Parasuraman et al., 2000).
Human-Automation Interaction 93
Figure 2.2. Stages of a typical large-scale system.
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 93
Levels of Automation
Table 2.1 is a scale of degrees of automation. Alternative forms of such a scale are dis-
cussed in the literature (Endsley & Kaber, 1999; Sheridan, 1992b; Sheridan & Verplank,
1978; Wei, Macwan, & Wierenga, 1998). Some versions scale the sensing (afferent) and
the motor (efferent) functions separately; some have more levels and some fewer. There
are two main points to be made: that automation need not be all or none—there are
various degrees appropriate to different problem contexts—and that different process
stages of a complex system are appropriately automated to different degrees.
Parasuraman et al. (2000) emphasized the latter point and gave examples of how
the appropriate level of automation differs at the four stages for different applications.
Endsley and Kaber (1999) demonstrated the level of automation effects on perform-
ance, situation awareness, and workload in a dynamic control task. Wei et al. (1998)
suggested a model for the appropriate degree of automation of different tasks based
on a task’s effect on system performance and its demand on the operator relative to
Criteria of Function Allocation and
Which functions should be relegated to the human and which to the automation is a
classical problem going back to the Fitts (1951/2005) MABA-MABA list (“Men are bet-
ter at ...;machines are better at . . .”). Although automation has progressed to the
point that the Fitts list is no longer valid, criteria for function allocation have not been
settled (Hancock & Scallen, 1996; Sheridan, 2000).
Human-centered automation is a phrase popularized by Billings (1997) and widely
used to convey that automation must be designed to work in conjunction with the
humans controlling or otherwise interacting with it; to engineer the automation and
expect the human to accommodate to it can be a recipe for disaster. By now the point
is well understood within the human factors community. But debate continues about
94 Reviews of Human Factors and Ergonomics
TABLE 2.1: A Scale of Degrees of Automation
1. The computer offers no assistance; the human must do it all.
2. The computer suggests alternative ways to do the task.
3. The computer selects one way to do the task and
4. executes that suggestion if the human approves, or
5. allows the human a restricted time to veto before automatic
6. executes the suggestion automatically, then necessarily informs the
7. executes the suggestion automatically, then informs the human only
8. The computer selects the method, executes the task, and ignores the
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 94
the criteria for whether automation is appropriately “human centered.” As suggested
by Table 2.2, the various criteria are different from one another, and for any one of
these criteria, serious questions can be raised about the extent of its applicability.
AUTOMATION-RELATED INCIDENTS AND ACCIDENTS
Much of the impetus for understanding human interaction with automated systems
stems from several accidents that have involved automation, an issue ﬁrst raised in the
aviation domain in a seminal paper by E. L. Wiener and Curry (1980). We focus on
incidents and accidents related to feedback about system states provided by the
automation, misunderstanding of automation, and overreliance on automation
(Billings, 1997; Parasuraman & Byrne, 2003; Parasuraman & Riley, 1997). These three
Human-Automation Interaction 95
TABLE 2.2: Some Criteria of Human-Centered Automation (and
Reasons to Question Them)
1. Allocate to the human the tasks best suited to the human, and
allocate to the automation the tasks best suited to it. (Unfortunately,
there is no consensus on how to do this; nor is the allocation policy
necessarily ﬁxed, but may depend on context.)
2. Keep the human operator in the decision-and-control loop. (This is
good only for intermediate-bandwidth tasks. The human is too slow
for high bandwidth and may fall asleep if bandwidth is too low.)
3. Maintain the human operator as the ﬁnal authority over the automa-
tion. (Humans are poor monitors, and in some decisions it is better
not to trust them; they are also poor decision makers when under
time pressure and in complex situations.)
4. Make the human operator’s job easier, more enjoyable, or more
satisfying through friendly automation. (Operator ease, enjoyment,
and satisfaction may be less important than system performance.)
5. Empower or enhance the human operator to the greatest extent
possible through automation. (Power corrupts.)
6. Support trust by the human operator. (The human may come to
overtrust the system.)
7. Give the operator computer-based advice about everything he or she
should want to know. (The amount and complexity of information is
likely to overwhelm the operator at exactly the worst time.)
8. Engineer the automation to reduce human error and minimize
response variability. (A built-in margin for human error and experi-
mentation helps the human learn and not become a robot; see
Rasmussen, Pedersen, & Goodstein, 1995.)
9. Make the operator a supervisor of subordinate automatic control
systems. (Sometimes straight manual control is better than supervi-
10. Achieve the best combination of human and automatic control,
where best is deﬁned by explicit system objectives. (Rarely does a
mathematical objective function exist.)
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 95
issues were also among the top 5 of more than 100 automation-related issues identi-
ﬁed in a survey of aviation experts by Funk et al. (1999), the other 2 issues being poor
display design and inadequate automation training.
Accident investigators have attempted to draw lessons learned from automation-
related accidents, many of which have been in aviation (Billings, 1997), although acci-
dents have also occurred in road, train, and maritime transportation; in manufactur-
ing and process control; and in health care. Commercial aviation has a very good safety
record, and modern highly automated aircraft are not only more fuel efﬁcient but also
safer than earlier generations of aircraft. Nevertheless, several highly publicized inci-
dents and accidents involving automated aircraft in the 1980s and 1990s, coupled with
the quest for even higher safety levels in the face of an increased volume of air traf-
ﬁc, have motivated greater scrutiny of automation by the aviation industry and the
FAA (Abbott et al., 1996).
Although potential problems have been identiﬁed in human interaction with use
of automation, it is not always easy to provide a succinct deﬁnition of an automation-
related incident or accident because of the multiplicity of precipitating events and
conditions ultimately leading to any accident (e.g., Reason, 1990; see also Wiegmann
& Shappell, 1997). Accident investigators have used the National Transportation
Safety Board Aviation Coding Manual (Aviation Coding Manual, 1998) and NASA’s
Aviation Safety Reporting System (Aviation Safety Reporting System, 2005), which
can be examined with appropriate keywords to identify automated-related incidents
(Funk et al., 1999).
Feedback on System States
Many so-called automation-related incidents occur even though the automation did
not malfunction.Rather,in many incidents,the state of the automated system changed,
but this was not communicated to the human operators in a salient way. An early
example from aviation is the 1972 crash of a Lockheed L-1011 in the Florida Ever-
glades. The ﬂight crew was engaged in troubleshooting a problem with a landing gear
indicator light and did not recognize that the altitude hold function of the autopilot
had been inadvertently disconnected (National Transportation Safety Board, 1973).
Although several factors contributed to this accident, a major one was poor feedback
on the state of automation provided by the system. The disengagement of automation
should be clearly signaled to the human operator so that it can be validated as intended
or unintended. Most current autopilots now provide an aural and/or visual alert upon
disconnect. The alert remains active for a few seconds or requires a second disconnect
command input by the pilot before it is silenced. Persistent warnings such as these,
especially when they require additional input from the pilot, are intended to decrease
the chance of an autopilot disconnect or failure’s going unnoticed.
However, a persistent warning may be insufﬁcient by itself if the crew does not
know what state the automation is in. As pointed out by Degani (2003) in his excel-
lent recent book Taming Hal, many automated systems—from simple alarm clocks to
automatic teller machines to large aircraft—involve internal transitions between dif-
ferent machine states or modes. Using state transition diagrams, Degani illustrated how
such transitions are sometimes hidden from the user, as a result of which the user may
96 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 96
think the machine is in one state when it is actually in another. With simple systems,
such as VCR/TV controls, this might lead only to annoyance or frustration as the user
fumbles with adjusting the TV set while the control is actually in VCR mode. But with
more complex systems, the lack of salient feedback about automation states (Norman,
1990) can lead to catastrophe.
Two decades after the L-1011 accident, an Airbus A300 experienced an in-ﬂight
incident off the coast of Florida (National Transportation Safety Board, 1998b). At the
start of a descent into the terminal area, the autothrottles were holding speed constant
but no longer controlled airspeed after the plane leveled off at an intermediate alti-
tude—a circumstance unknown to the pilots. The aircraft slowed gradually to almost
40 knots below the last airspeed set by the pilots and subsequently experienced stalling
after the stall warning activated. There was no evidence of autothrottle malfunction.
Rather, the crew apparently believed that the automated system was controlling air-
speed when in fact it had disengaged, which could be done with a single press of the
disconnect button. When the system was disengaged, the green mode annunciator in
the primary ﬂight display would change to amber, and the illuminated button on the
glare shield used to engage the system would be extinguished.
The NTSB (1998b) noted that although the change in the annunciators could serve
as a warning, the format of the displays did not command attention because they were
passive and persistent. The NTSB also pointed to autothrottle disconnect warning sys-
tems in other aircraft that required positive crew action to silence or turn off. These
systems incorporated ﬂashing displays and, in some cases, aural alerts that would help
capture the pilot’s attention in the case of an inadvertent disconnect. These systems
more rigorously adhere to the principle of providing salient feedback to the operator
about the state of an automated system.
Misunderstanding or Lack of Understanding of Automation
A major characteristic of automated systems is complexity. Designers and engineers
have developed automated systems with such large numbers of interacting subcom-
ponents that understanding the effects of all possible interactions has become increas-
ingly difﬁcult, if not impossible. Much previous research has focused on misunder-
standings of complex automated systems in the cockpit, such as the ﬂight management
system (FMS), whose complexity is so great that even highly skilled and trained oper-
ators such as commercial pilots can have difﬁculty understanding the nuances of its
behavior (Sarter & Woods, 1995).
It has been suggested that misunderstandings arise because of a mismatch between
the mental model of the pilot and the behaviors of the automated system as pro-
grammed by the designers (Sherry & Polson, 1999). Several examples of incidents and
accidents resulting from these system misunderstandings have been reported (Billings,
1997; Funk et al., 1999; Sarter & Woods, 1995). Although some have had benign out-
comes and become lessons learned, others have involved serious loss of life.
For example, in 1994, an A300 crashed in Nagoya, Japan, after the pilots inadver-
tently engaged the autopilot’s go-around mode. The pilots countered the unexpected
pitch-up by making manual inputs, which turned out to be ineffective (Billings, 1997).
Essentially, the pilot attempted to continue the approach by manually deﬂecting the
Human-Automation Interaction 97
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 97
control column, which in all other aircraft—and in this aircraft in all modes except
the approach mode—would normally disconnect the autopilot. However, in this par-
ticular aircraft and in this particular mode, the autopilot had to be manually dese-
lected and could not be overridden by control column inputs. Consequently, a power
struggle developed between the pilot and the autopilot, with the pilot attempting to
push the nose down through elevator control and the autopilot attempting to lift the
nose up through trim control. This caused the aircraft to become so far out of trim
that it could no longer be controlled.
Overreliance on Automation
Automated systems typically are highly reliable—with the exception of some auto-
mated alerting systems, which can have high false alarm rates. This, together with their
opacity and complexity,can lead operators to rely unquestioningly on automation.The
phenomenon of overreliance on automation has been likened by Mosier, Skitka,Heers,
and Burdick (1998) to a decision bias, or automation bias. They suggested that the
bias is reﬂected in the operator’s following the advice of automated systems even when
the automation commits both errors of omission (misses) and errors of commission
However, reliance on automation can be distinguished from compliance. Meyer
(2001) showed that when automation reliability is such that malfunctions are almost
always correctly indicated, the automation makes few misses, so operators have high
reliance on the automation. This is an effective strategy but can result in a problem
when the automation does fail to indicate a hazard (a miss), because in this case the
operator may not monitor the automation—the so-called complacency effect (Para-
suraman et al., 1993). On the other hand, if automation reliability is such that few
false alarms are made, then the operator usually has high compliance: If an automated
alarm sounds, the operator tends to comply immediately with the alarm and attend
to the situation. Reliance on automated aids permits the operator to attend to tasks
other than the automated task until the alert is triggered, thus improving multitask
performance and not just the performance on the automated task.
An example from the maritime industry is particularly revealing of the effects of
user overreliance on automated systems. The cruise ship Royal Majesty ran aground
off Nantucket after veering several miles off course into shallow waters. Fortunately,
there were no injuries or fatalities as a result of the accident, but losses totaled $2 mil-
lion in structural damage and $5 million in lost revenue. The automated systems in
this ship included an autopilot and an Automatic Radar Plotting Aid that was tied to
signals received by a Global Positioning System (GPS). Under normal operating con-
ditions, the autopilot used GPS signals to keep the ship on its intended course. How-
ever, the GPS signals were lost when the cable from the antenna frayed (it was placed
in an area of the ship where many sailors walked). As a result, the GPS and autopilot
switched to dead reckoning mode, no longer correcting for winds and tides, which in
this case carried the ship toward the shore.
According to the NTSB report on the accident, the probable cause was the crew’s
overreliance on the Automatic Radar Plotting Aid and managers’ failure to ensure that
crewmembers were adequately trained in understanding the automation features and
98 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 98
its capabilities and limitations. The report went on to state, “the watch ofﬁcers’ mon-
itoring of the status of the vessel’s GPS was deﬁcient throughout the voyage,” and “all
the watch-standing ofﬁcers were overly reliant on the automated position display . . .
and were, for all intents and purposes, sailing the map display instead of using navi-
gation aids or lookout information.”
This accident represents a classic case of automation complacency related to inap-
propriately high trust in the automation (Lee & See,2004; see Degani, 2003,for a more
detailed account of the accident). This accident also demonstrates the importance of
salient feedback about automation states and actions, as mentioned earlier. The text
annunciators that distinguished between the dead reckoning and satellite modes were
not salient enough to draw the crew’s attention to the problem.
A general aviation accident further exempliﬁes the danger of overreliance on auto-
mated systems such as GPS. In 1997, a single-engine airplane with a non-instrument-
rated pilot took off under instrument meteorological conditions (National Trans-
portation Safety Board, 1998a). About two hours later, following a meandering course
that included course reversals and turns of more than 360°, the aircraft crashed into
trees at the top of a ridge. No mechanical problems with the airplane’s controls,engine,
or ﬂight instruments were identiﬁed. A person who spoke with the pilot before depar-
ture stated that the pilot “was anxious to get going. He felt he could get above the
clouds. His GPS was working and he said as long as he kept the [attitude indicator]
steady he’d be all right. He really felt he was going to get above the clouds.” Many fac-
tors undoubtedly played a role in this accident, but the apparent reliance on GPS tech-
nology, perhaps to compensate for insufﬁcient training and lack of ratings, stands out
as a compelling factor.
There is some consensus for the existence of overreliance on automation (called
complacency) for now-entrenched historical reasons related to the development of
NASA’s Aviation Safety Reporting System (see Billings, Lauber, Funkhouser, Lyman, &
Huff, 1976). However, Moray and Inagaki (2001) disagreed with the interpretation of
Parasuraman et al. (1993) that their ﬁndings reﬂected complacency. Though accepting
that people have often failed to notice when automation fails, they argued that none
of the reported literature has shown what the optimal level of trust would actually be.
Consistent with the information-theoretic model of sampling (Senders, 1964), Moray
and Inagaki (2001) proposed that automation should be monitored (or sampled) by
the human at a rate set by the objective failure rate of the automation: The more reli-
able the automation, the less the operator should monitor it. Moray and Inagaki stated
that complacency (overtrust) should be inferred only if the operator sampled less fre-
quently than this rate; if they sampled more frequently, they should be characterized
as “skeptical” (undertrusting).
This is an elegant theory that has some support. Moray et al. (2000) found that in
an adaptive automation experiment with varying levels of automation, participants
converged on the optimal level of trust—in the sense that subjective trust matched the
objective reliability of the automation—relative to higher or lower trust. Moray and
Inagaki’s (2001) theory depends for its validation on one’s being able to specify pre-
cisely what the optimal sampling rate of automation should be. This is easy when the
failure rate of the automation is known. But monitoring automation is rarely the only
task for the operator, and therefore the sampling rate also depends on the operator’s
Human-Automation Interaction 99
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 99
being able to specify the required sampling frequency of all the other tasks in which
he or she is engaged. In general, identifying an optimal automation sampling rate may
be difﬁcult to do in complex systems with multifunction forms of automation and
simultaneous manual tasks. Moreover, the theory cannot explain why operators appar-
ently sample less frequently when the automation failure rate remains constant than
when it varies at about the same mean rate.
Parasuraman et al. (1993) found that the so-called complacency effect was consid-
erably reduced when automation reliability was variable over time compared with
when it was constant (a ﬁnding recently replicated by Bagheri & Jamieson, 2004).
The overall failure rate was the same in both conditions, but operators noticed
more automation failures in the variable-reliability condition than in the constant-
reliability condition, perhaps because they were more skeptical.
One explanation for undermonitoring of automation that complements the trust
theory is based on attention. Common sense tells us that there is nothing to be gained
by attending to very reliable automation with low downside risk—for example, our
home heating systems—until after failure is evident. Attending to imperfect automa-
tion is also diminished when the operator is engaged in other tasks that require focal
attention.Evidence in support of this view comes from studies of eye movements. Met-
zger and Parasuraman (2001), for example, asked experienced air trafﬁc controllers to
monitor a radar display for separation conﬂicts while simultaneously accepting and
handing off aircraft to and from their sector, managing electronic ﬂight strips, and
using data linking to communicate with pilots. They were assisted by a “conﬂict probe”
aid that predicted the future (up to 8 minutes) courses of pairs of aircraft in the sec-
tor. The automation was highly reliable and reduced the time that controllers took to
call out the conﬂict (see also Metzger & Parasuraman, 2005).
In one scenario, however, the automation did not point out the conﬂict because it did
not have access to the pilot’s intent to change course. Controllers were either consider-
ably delayed or missed the conﬂict entirely. Eye movement analysis showed that those
controllers who did not detect the conﬂict in this case had fewer ﬁxations of the radar
display compared with when they had been given the same conﬂict scenario without the
conﬂict probe aid. This ﬁnding is consistent with the view that overreliance on automa-
tion is associated with reduced attention allocation, compared with manual conditions.
HUMAN PERFORMANCE RESEARCH
RELATED TO AUTOMATION
Trust as a Design and Performance
Issue in Human-Automation Interaction
Until recently system design engineers seldom—if ever—used the term trust when dis-
cussing automation, but today it is seen as a key concern (Lee & See, 2004; Parasura-
man & Riley, 1997).
For different people and contexts the term can have different meanings (Sheridan,
1988). Trust can be both a cause and an effect. It is a cause in the sense that a human’s
100 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 100
use of the automation depends on his or her trust of it. As an effect, it can have sev-
1. judged reliability of the automation, in the usual sense of repeated, consistent function-
2. perceived robustness—that is, demonstrated or promised ability to perform under a variety
of circumstances; sense of familiarity, whereby the system employs procedures, terms, and
cultural norms that are familiar; perceived understandability, in the sense that the human
supervisor can form a mental model and predict future system behavior; usefulness of the
system to the trusting person; or dependence of the trusting person on the system.
Lee and See (2004) pointed out that user or system vulnerability to automation
error is a critical component of the deﬁnition of trust; something important must be
riding on the decision to rely on the automation. This is an important point because
in numerous recent “automation” studies, the experimental participants have nothing
at stake if the automation fails. In prior work, Riley (1996) demonstrated that per-
ceived risk has a signiﬁcant inﬂuence on automation reliance. His concern was that
some researchers may be losing sight of the fact that automation failures in the real
world can have real consequences, that human biases regarding errors of omission and
commission are related to the nature of the real potential outcomes, that these biases
will likely apply to automation use decisions, and therefore that studies that do not
incorporate the element of operator vulnerability to automation error may yield
potentially misleading results (see also Parasuraman & Riley, 1997).
Lee and Moray (1992) studied trust during the supervisory control of a simulated
pasteurizing plant in which operators could select either manual or automatic control
in the face of random failures.They proposed a quantitative model of dynamic changes
in operators’ trust and their decisions to use the automation, in which trust depended
on the past level of trust, failure probability, and subjectively perceived capability in
manual control. Shifts between automatic and manual control modes could be pre-
dicted by the ratio between trust in the machine and self-conﬁdence in one’s own man-
ual performance (Lee & Moray, 1992, 1994; Muir, 1988).
An interesting follow-up study was carried out by Lewandowsky, Mundy, and Tan
(2000) in which people shared control either with a human or with automation but
were told in all cases that they were sharing with a human. They were more tolerant
of system errors when they believed it was a human.
Engendering trust is often a desirable feature of a system, something the designer
strives for—but not always. Too much trust (usually naive trust) can be just as bad as
too little. Parasuraman and Riley (1997) made a compelling case that system design-
ers should be concerned about misuse, disuse, and abuse of automation based on dis-
trust and overtrust as well as on workload and other factors.
“Etiquette” in the Design of
Human-Automation Interactive Software
In their recent review of the literature on user trust in automation, Lee and See (2004)
proposed that trust is related to emotions and attitudes that people have regarding
automation. As Nass and colleagues (Nass, Moon, Fogg, Reeves, & Dryer, 1995; Reeves
Human-Automation Interaction 101
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 101
& Nass, 1996) showed, people often respond socially to computers in ways similar to
how they interact socially with other people. Computers are becoming more intelli-
gent—more like people, some might say. Because machines and the people who use
them need to communicate and cooperate, and because people are accustomed to
communicating with other people, it has become clear that the same social mores that
apply to people might also apply to intelligent machines.
Individuals are typically most attracted to others who appear to have personalities
similar to their own. This phenomenon, which psychologists call the social attraction
hypothesis, also predicts user acceptance of computer software (Nass et al.,1995). These
considerations suggest that etiquette, or adherence to an accepted but frequently
implicit code of behavior between individuals in any social setting, may also play a key
role in human-computer relations.
What is commonly considered to be good etiquette is a set of behavioral practices
that make interaction between mature people acceptable and efﬁcient (Grice, 1975). If
this were not so, etiquette might have died out long ago. Grice posed four maxims for
cooperation in conversation (etiquette):
1. Maxim of quantity: Say what serves the present purpose but not more.
2. Maxim of quality: Say what you know to be true based on sufﬁcient evidence.
3. Maxim of relation: Be relevant, to advance the current conversation.
4. Maxim of manner: Avoid obscurity of expression, wordiness, ambiguity, and disorder.
Grice’s axioms were considered by C. A. Miller (2004) in the context of designing
adaptive user interfaces. His proposed rules, slightly abbreviated, are as follows:
1.Make many conversational moves for every error made.
2.Make it very easy to override and correct any errors.
3.Know when you are wrong, mostly by letting the human tell you.
4.Don’t make the same mistake twice.
5.Don’t show off. Just because you can do something does not mean you should.
6.Talk explicitly about what you are doing and why. (Your human counterparts spend a lot
of time in such meta-communication.)
7.Use multiple modalities and information channels redundantly.
8.Don’t assume every user is the same; be sensitive and adapt to individual, cultural, social,
and contextual differences.
9.Be aware of what your user knows, especially what you just conveyed (i.e., don’t repeat
10.Be cute only to the extent that it furthers your conversational goals.
The work of C. A. Miller (2004) suggests that automation that follows such agreed-
upon axioms of etiquette is more likely to be accepted and liked by human operators.
But is there any objective evidence that designing etiquette into automation can improve
system performance? In other words, does etiquette matter as far as the bottom line is
concerned—productivity, efﬁciency, safety—or is it simply something that makes the
operator feel good, an adjunct to the principle of human-centered automation?
A recent experiment by Parasuraman and Miller (2004) suggests that etiquette can
inﬂuence efﬁciency and therefore possibly safety. These authors examined whether
etiquette affects human decisions to use automation effectively in high-workload
102 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 102
situations and, if so, whether in the same way as automation reliability inﬂuences user
trust and usage. Unreliable or imperfect automation is generally correlated with
decreased trust and decreased reliance on automation, but various other factors mod-
erate this phenomenon (Lee & See, 2004). To what extent is etiquette one such factor?
Parasuraman and Miller investigated whether good automation etiquette could
compensate for poor reliability and result in increased use of automation, and, con-
versely, whether poor etiquette could negate the beneﬁts of high reliability and result
in decreased trust and automation usage. Such phenomena are not uncommon in
human-human relationships. To test the hypothesis, 16 participants (general aviation
pilots and nonpilots) were examined on a ﬂight simulation task, the Multi-Attribute
Task (MAT) Battery (Comstock & Arnegard, 1992), which incorporates a primary
ﬂight (tracking) task, a fuel management task, and an engine health monitoring task.
Participants performed the ﬁrst two tasks manually at all times so as to simulate a
Intelligent automation support modeled after the Engine Indicator and Crew Alert-
ing System (EICAS) that is typically installed in modern automated aircraft was pro-
vided for the engine systems monitoring task. In this task, participants had to check
particular engine parameters for malfunctions and make an appropriate diagnosis. They
were asked to query the EICAS system while using the automation support tool, which
provided advice on possible malfunctions.An example advisory message would be,“The
Engine Pressure Ratio (EPR) is approaching Yellow Zone. Please check. Also, cross-
check Exhaust Gas Temperature (EGT). There is a possible ﬂame-out of Engine 1.”
Parasuraman and Miller (2004) deﬁned good automation etiquette as a communi-
cation style that was “non-interruptive” and/or “patient.” Conversely, poor automation
etiquette occurred when the automation communicated in an “interruptive” or “impa-
tient” manner. For example, in the interruptive (poor etiquette) case, the automation
provided advice without warning and came on when the user was already querying
EICAS and was engaged in fault diagnosis (i.e., already engaged in the behavior the
system was recommending). In the impatient case, the automation urged the next
query before the user was ﬁnished with the current query. (The experimenters con-
ﬁrmed their deﬁnitions of good and poor etiquette by questioning the participants at
the end of the experiment.)
The good and poor etiquette interfaces were combined with two levels of automa-
tion reliability, following a previous study by Parasuraman et al. (1993). In the high-
reliability condition, the EICAS provided correct advice on 8 of 10 engine malfunctions
(80%), whereas it was correct on only 6 of 10 malfunctions (60%) in the low-reliability
condition. The results were, as expected, that user diagnostic performance was better
when the reliability of the automation was high (80%) than when it was low (60%).
However, and less obviously, good automation etiquette signiﬁcantly enhanced diag-
nostic performance both when automation reliability was low and when it was high.
Third, and perhaps most interestingly, the effects of automation etiquette were power-
ful enough to overcome low automation reliability: Performance in the low-reliability/
good-etiquette condition was almost as good as (and not signiﬁcantly different from)
that in the high-reliability/poor-etiquette condition. These performance ﬁndings were
paralleled by similar effects on user ratings of trust in the automation.
Human-Automation Interaction 103
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 103
Finally, the effects of poor etiquette on human-system performance were not
attributed solely to distraction or interruption. All messages, whether associated with
good or poor etiquette, were presented visually in the EICAS communications win-
dow and not, for example, using speech synthesis, which could have been distracting.
Furthermore, a control group of participants was run to test the hypothesis that any
interruption might be expected to degrade user performance. These participants
received nonspeciﬁc interruptions—for example, “Maintaining primary ﬂight per-
formance is important, but do not forget to check engine parameters for possible mal-
function.”These neutral interruptions had no effect compared with the good and bad
These results provide strong evidence for the inﬂuence of automation etiquette on
both user performance and trust in using an intelligent fault management system to
diagnose engine malfunctions, at least in a high-workload setting when users are busy
doing other tasks. The results also clearly show that reliability per se may not be suf-
ﬁcient to promote overall human-machine system efﬁciency: Both user diagnostic per-
formance and trust were lowered by poor automation etiquette even when the relia-
bility of the advice provided by the automation was relatively high.
The results also suggest the intriguing notion that good automation etiquette can
compensate for the performance costs associated with low automation reliability. Some
may ﬁnd this result disturbing, for it suggests that developing robust, sensitive, and
accurate algorithms for automation—a challenging task under the best of circum-
stances—may not be necessary so long as the automation “puts on a nice face” for the
user. This is unlikely, however, because this experiment also clearly showed that the best
user performance (and the highest trust) was obtained in the high-reliability condition
in which the automation also communicated its advice to the user with good etiquette.
Ecological Interface Design and Automation
Systems that provide the user with information concerning automation modes, system
states, and future automated actions—particularly if they do so with good etiquette—
can improve human—automation communication and therefore potentially enhance
system performance.If the provision of such feedback requires extensive cognitive pro-
cessing on the part of the user, however, any beneﬁt may be counteracted by the
increased cognitive load on the user. Consequently, there is a need to develop inter-
faces that provide feedback on automation states and behaviors in a manner that
requires little or no cognitive effort but can be directly apprehended by a quick glance
at a display that provides the appropriate avenue for rapid action. Such interfaces have
collectively been characterized under the rubric of ecological interface design (EID;
Vicente, 2002; Vicente & Rasmussen, 1992). The origin of the term ecological can be
traced to the perceptual theories of Gibson (1979), whose concepts of affordance and
direct perception are closely paralleled by EID.
The EID approach also draws on the concept of the goals-means-ends abstraction
hierarchy proposed by Rasmussen (1986) and on his taxonomy of skills, rules, and
knowledge. In this approach a given domain of work is decomposed into parts, begin-
ning with the top goal, then into the means by which the goal is achieved, both in
104 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 104
terms of abstract functions and, at the bottom of the hierarchy, in terms of the phys-
ical acts needed to achieve those functions. In theory, EID interfaces are thought to
allow for direct perception of functional relationships between system components,
without the need for extensive cognitive processing, and to enable rapid action. In
essence, the EID interface is thought to replace cognition with perception, thereby
facilitating action. Accordingly, human-automation interaction might also be facili-
tated if the interfaces used to provide feedback to the user on automation states are
designed in accordance with EID principles.
Two studies reported by Furukawa, Parasuraman, and Inagaki (2003) are described
for the purpose of illustrating the efﬁcacy of EID interfaces for human-automation
interaction. In the ﬁrst study (see also Molloy & Parasuraman, 1994), an integrated
display with an emergent perceptual feature was used to examine pilot fault manage-
ment performance in a ﬂight simulation task. As discussed previously, when operators
are busy with several tasks, they may fail to monitor the status of an automated task,
so that when occasional failures occur, operators are less likely to detect the failure (or
are slow in responding) compared with when they perform the task manually (Para-
suraman et al., 1993). This automation complacency effect reﬂects a policy of allocat-
ing attention away from the automated task to the other tasks that the operator has
to perform simultaneously (Metzger & Parasuraman, 2005; Moray & Inagaki, 2001;
Parasuraman et al., 1993). Hence, one way to mitigate the effect would be to use an
integrated display (Bennett & Flach, 1992) based on EID principles to display system
states, both under normal conditions and when malfunctions occur.
In the ﬁrst study reported by Furukawa et al. (2003), 12 general aviation pilots per-
formed a ﬂight simulation task involving the following subtasks: two-dimensional
compensatory tracking (a resource management task requiring balancing of the fuel
tanks of the aircraft) and an automated engine systems failure detection task. The
interface used to display the engine systems task consisted of either an integrated or a
nonintegrated display. The pilot’s task was to detect deviations in different engine
parameters (EPR, N1, EGT, etc.) and take appropriate corrective actions.
The nonintegrated display was based on a traditional EICAS, which was displayed
with circular gauges showing engine parameters in an analog display. The integrated
display was based on the Engine Monitoring and Crew Alerting System (EMACS),
which was depicted with a deviation bar graph. The tips of the four bars displayed an
emergent perceptual feature—an implied contour line when the state of the engine
was normal—which provided a ready indication of system normality or malfunction
without the expenditure of much cognitive effort. In Rasmussen’s (1986) terms, the
line provided the abstract information obtained from the state parameters during nor-
mal operation. If one of the engines failed, the emergent line was distorted, enabling
the pilots to detect it easily.
There were no differences in pilot performance between the EICAS and EMACS
displays under normal conditions. However, when the automation failed to detect and
diagnose malfunctions, pilot monitoring performance (detection and diagnosis rates)
was signiﬁcantly better for the integrated EMACS display than for the nonintegrated
EICAS display. Moreover, the automation complacency effect—the reduction in per-
formance for automated compared with manual monitoring—was eliminated with the
Human-Automation Interaction 105
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 105
integrated display. Analysis of eye movements indicated that dwell times on the dis-
play were signiﬁcantly shorter with the EMACS than with the EICAS display.
These ﬁndings indicate that automation complacency results from a policy in which
attention is allocated away from the automated task to the other manual tasks that the
operator has to perform. As a result, if abnormalities in the automated control of a
task occur, malfunctions may not be rapidly detected and diagnosed. However, if these
are shown to the operator using an integrated display based on EID, fewer attentional
resources are needed to detect and diagnose the malfunction. This was supported by
the eye movement analysis: Pilots had shorter visual ﬁxations (dwells) with the EMACS
display, yet they performed better than with the EICAS display, suggesting that display
integration facilitated efﬁcient information extraction. Furthermore, EID eliminated
the automation complacency effect.
In a second experiment, Furukawa et al. (2003) used a modiﬁcation of the DURESS
process control simulation (Vicente & Rasmussen, 1992) to which automated con-
trollers with three different modes were added. As a result, participants had consider-
able difﬁculty in recognizing automation states and changes using a standard display
of the process control simulation.EID was used to design a new display that supported
operators in their ability to prevent conﬂicts with the automated controllers by explic-
itly representing the intentions (i.e., goals and means) of the automation on the dis-
play. Performance in 16 participants was compared with the intention-represented EID
and a standard EID display. Participants were shown one of the displays for 8–10 sec-
onds and then were asked to report key parameters in the current state of the system
as well as its future state. They also had to indicate the control operations that needed
to be taken and the target states of the operations. The scenarios included situations
that had been experienced before as well as unexpected events.
The results showed that with such a mode-rich, high-capability automated system,
operators found it difﬁcult to recognize the goals, the means, and their interrelations
(means-end relations) of the automated controllers using the standard display. Some
participants failed to supervise particular tasks performed by the automated con-
trollers, perhaps because they thought that controllers had complete responsibility for
the tasks, and even though they clearly demonstrated their comprehension of the sit-
uation. However, performance was enhanced when the intention-represented EID dis-
play was used. With this display, the automated controllers revealed their goals and
means to participants, like human supervisors or coworkers usually do.As a result, par-
ticipants could recognize the means-end relations through observing the behaviors of
the automated controllers. Consequently, they took appropriate actions that were not
in conﬂict with those of the automatic controllers, thereby enhancing performance.
In general, these results indicate that EID displays might be particularly helpful for
human-automation interaction under nonroutine conditions. Under routine operat-
ing conditions that have been experienced many times before, human operators can
effectively monitor and control a complex system using standard operating procedures.
Although this is efﬁcient for normal operations, it may not be appropriate for abnormal
or unanticipated conditions. When unexpected events occur, a human-automation
display interface is required that allows for quick comprehension of system state and
106 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 106
Quantitative Models for Supervisory Functions
As discussed previously, human-automation interaction can require several different
functions served by the human supervisor and/or the automation, as shown by the
boxes at the top of Figure 2.1. Quantitative models for each of these supervisor func-
tions are available in the literature (see Table 2.3).
Most of these models were developed by engineers for engineering applications that
had little to do with human interaction (and mostly in conjunction with World War
II). Since 1950, many such engineering theories have been adapted to the analysis and
modeling of human behavior. For example, following Shannon’s exposition of infor-
mation theory (1947) applied to telephone communication systems, psychologists
modeled a variety of sensory-motor skill tasks (Hick, 1952) using an information
transmission model (bits per second), and G. A. Miller (1956) applied information
transmission (in bits per selection) to performance in immediate memory tasks.
Following the development of signal detection theory for sonar and radar detec-
tion applications, Green and Swets (1966) and Swets (1996) applied that theory to sen-
sory psychophysics. And following the development of classical automatic control the-
ory for gun control (James,Nichols,& Phillips,1947; Craik,1947), McRuer and Krendel
Human-Automation Interaction 107
TABLE 2.3: Theoretical Models Relevant to the Five Supervisory
Function Appropriate Type of
Model physical system System analysis/differential
Satisﬁce trade-offs among objective Multiattribute utility theory
Formulate control strategy Control theory, optimization theory
Information and coding theories
Allocate attention Sampling theory, Markov network
Estimate state Estimation theory
Detect failure Signal detection, Bayesian analysis,
pattern recognition theories
Human and machine learning
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 107
(1959) and McRuer and Jex (1967) applied the control-loop model to human pilot
control of aircraft in turbulence. Kleinman, Baron, and Levison (1970) applied the
more recent optimal control theory to pilot instrument scanning and control of an
aircraft during approach.
These and other adaptations of engineering theories to simple sensorimotor skills
have been reviewed by various authors (Rouse, 1980; Sheridan & Ferrell, 1974), but
none of the aforementioned modeling efforts really dealt with human-automation
interaction. Many classical papers (including the classical control papers mentioned
earlier) were recently reprinted in Moray (2005).
Understanding the physical interactions of the processes to be controlled, setting of
goals and constraints, and devising a control strategy (all part of planning); commu-
nicating the model, the goals, constraints, and control strategy to a computer (teach-
ing); allocating attention and estimating system state (monitoring); preparing for
emergency action (intervening); and learning from experience (learning) are all rela-
tively high level cognitive tasks—and we really do not have an acceptably robust quan-
titative model for even one of these functions, let alone all of them working together.
So,in spite of the seeming relevance of the above-listed engineering models to the var-
ious human supervisory functions, the challenge remains to make the relevance
explicit and to devise tractable predictive models.
Comparing Performance Attributes of
Humans and Automated Systems
One type of model that is relatively easy to realize is that in which certain character-
istics of an automatic system can be speciﬁed and comparisons made with either a
human or another automatic system in terms of those characteristics. For example,
Sheridan and Parasuraman (2000) offered a simple analytical criterion for deciding
whether a human or automatic system is better in a failure detection task. The method
is based on expected-value decision theory in much the same way as is signal detec-
tion. It requires speciﬁcation of the probabilities of misses (false negatives) and false
alarms (false positives) for both the human and the automation being considered, as
well as factors independent of the choice—namely, costs and beneﬁts of incorrect and
correct decisions and the prior probability of failure. The method can also serve as a
basis for comparing different modes of automation. The authors discuss some limit-
ing cases of application as well as some decision criteria other than expected value.
Thus far we have discussed the effects on system performance of various forms of
automation, whether deﬁned by different levels of autonomy or by different stages of
information processing to which machine aiding can be applied. One goal of much of
this research is to inﬂuence system design by recommending a particular level or type
of automation or, as in the model by Sheridan and Parasuraman (2000), deciding
whether to automate at all. As such, this work assumes that once the ﬁnal design has
108 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 108
been identiﬁed, the automation has a ﬁxed level or type consistent with these design
features.This approach,in which the characteristics of automation are set at the design
stage and then executed in the same way during operational use, has been referred to
as static automation (Parasuraman, Bahri, Deaton, Morrison, & Barnes, 1992). In con-
trast, in adaptive automation, the level and/or type of automation is not ﬁxed but may
change during system operations.
Adaptive automation is akin to dynamic function allocation, in which the division
of labor between human and machine agents is not ﬁxed but dynamic, ﬂexible, and
context dependent (Hancock, Chignell, & Lowenthal, 1985; Kaber & Endsley, 2004;
Parasuraman, 1993; Parasuraman et al., 1992; Rouse, 1988; Scerbo, 1996, 2001; Ina-
gaki, 2003). For example, if performance in a higher level of automation is getting
worse, the automation may change to a lower level and/or turn over more or even all
control to the human. If, on the other hand, high human workload is detected and/or
the human is not responding appropriately, automation may go to a higher level so as
to become less dependent on the human.
The thorny human factors issue of allocation of function has typically been based
on stereotypical characteristics of human and computer capabilities, an approach that
has met with marginal success (Dialogs on Function Allocation, 2000; Jordan, 1963;
Sheridan, 1998). Hancock et al. (1985) suggested that the issue and its attendant prob-
lems may be bypassed if function allocation is viewed as dynamic rather than static.
In this view, allocation of function is not just a design activity but something that
occurs during system operations. Accordingly, there has been interest in investigating
the effects of different forms of adaptive automation on human performance in vari-
ous simulated tasks.
Techniques for Adaptive Automation
Although the adaptive automation concept is not new (Rouse, 1976), technologies for
its effective implementation have been explored only recently. Key technology devel-
opment efforts include the Air Force’s Pilot’s Associate (Hammer & Small, 1995), the
Navy’s Adaptive Function Allocation for Intelligent Cockpits (Morrison & Gluckman,
1994; Parasuraman et al., 1992), and the Army’s Rotocraft Pilot’s Associate programs
(C. A. Miller & Hannen, 1999). At the same time, several researchers have investigated
human performance in relation to adaptive automation using simulations of ﬂight, air
trafﬁc control, driving tasks, and process control (Inagaki, 2003; Moray et al., 2000;
Parasuraman, 1993; Scerbo, 1996).
A fundamental issue in all these areas concerns the means by which adaptive
automation is invoked. In an early review, Parasuraman et al. (1992) identiﬁed ﬁve
main categories of techniques for implementing adaptive automation: critical events,
operator performance measurement,operator physiological assessment, modeling, and
hybrid methods combining one or more of these techniques.
In the critical events method,automation is invoked if certain external events occur,
but not otherwise. For example, Barnes and Grossman (1985) described an aircraft air
defense system in which the beginning of a “pop-up” weapon delivery sequence led to
the automation of all defensive measures of the aircraft. However, if this critical event
Human-Automation Interaction 109
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 109
did not occur, the automation was not invoked. There are many other applications in
which the occurrence of critical events triggers automation of some kind.
The critical events method is ﬂexible in that it can be tied to current tactics and
doctrine during mission planning. A disadvantage of the method is its possible insen-
sitivity to actual system and human operator performance. For example, this method
will invoke automation irrespective of whether or not the pilot requires it when the
critical event occurs.
Measurement of operator performance and physiological measurement may be used
to overcome this limitation. Here the goal is to change the level or type of automation
based on an assessment of operator states. Various operator states (e.g., mental work-
load, fatigue, or—more ambitiously—operator intentions) may be inferred on the basis
of performance or other measures and the resulting value input into the adaptive logic
of an expert system or neural network. Kaber and Riley (1999), for example, used a
secondary-task measurement technique to assess operator workload in a target acqui-
sition task (see also Kaber & Endsley, 2004). They found that adaptive computer aiding
based on the secondary-task measure enhanced performance on the primary task.
Operator physiological assessment offers another potential input for adaptive sys-
tems (Byrne & Parasuraman, 1996; Parasuraman et al., 1992). For example, physio-
logical measurements may indicate that a human operator is dangerously fatigued or
experiencing extremely high workload. An adaptive system could use these measure-
ments to provide computer support or advice to the operator that would mitigate the
potential danger. Technology is available to measure a number of physiological signals
from the operator, from autonomic measures such as changes in heart rate to central
nervous system measures such as the EEG and event-related potentials (ERPs), as well
as measures such as eye scanning and ﬁxations (see Scerbo et al., 2001, for a review).
The main advantage of (at least some) physiological measures is their high band-
width compared with most performance measures (with the exception of performance
in manual tracking). There is now a substantial literature indicating that several psy-
chophysiological measures can be used for real-time assessment of mental workload
(Scerbo et al., 2001). Prinzel, Freeman, Scerbo, Mikulka, and Pope (2000) also specif-
ically demonstrated the feasibility of an adaptive system based on EEG measures.
Wilson and Russell (2003) used an artiﬁcial neural network (ANN) with multiple
physiological measures to identify low- and high-workload periods in real time dur-
ing performance of the MAT ﬂight simulation. When the ANN detected a high work-
load state, two of the MAT subtasks were automated. Wilson and Russell (2003) found
that implementing adaptive automation with the ANN method led to improved per-
formance on the MAT subtasks compared with manual performance. However, one
limitation of this study is that no comparison to static automation was made, so that
the speciﬁc beneﬁt of adaptive automation was not identiﬁed.
The four major techniques for implementing adaptive automation have comple-
mentary merits. The critical event technique has the advantage that it can be tied to
mission planning but the disadvantage that it does not take into account operator
requirements. Measurement of operator performance or physiological state has the
advantage of being potentially responsive to unpredictable changes in human operator
cognitive states. Physiological measures can be designed to be relatively unobtrusive,
but their sensitivity and validity need to be established in each application domain.
110 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 110
Moreover, all operator assessment methods are only as good as the sensitivity and
diagnosticity of the measurement technology. Modeling techniques have the advantage
that they can be implemented ofﬂine and easily incorporated into rule-based expert
systems (Parasuraman et al., 1992). However, this method requires a valid model, and
many models may be required to deal with all aspects of human operator perform-
ance in complex task environments. Furthermore, different models might give con-
trary decisions at a particular moment.
For these reasons, therefore, hybrid methods that attempt to optimize the relative
beneﬁts and disadvantages of each of these techniques may offer the best general
approach to implementing adaptive automation. Many investigators are attempting to
develop such hybrid approaches to adaptive automation (e.g., St. John, Kobus, Morri-
son, & Schmorrow, 2004), but nothing practical has been produced to date.
Mitigation of Driving Distraction by
Adaptive Information Systems
An active current research area in adaptive automation is to make information tech-
nology in highway vehicles adaptive to trafﬁc and other workload demands (Lee,
Caven, Haake, & Brown, 2001; Lee, McGehee, Brown, & Reyes, 2002). There is con-
cern that drivers are being distracted from the driving task by use of cell phones, radio
and other entertainment systems, navigation systems, personal digital assistants
(PDAs), and, possibly in the future, e-mail, faxes, stock trading, and all sorts of real-
time interactions with personal or corporate computer ﬁles. The highway vehicle can
easily become the ofﬁce. The problem is one of maintaining safety.
If distraction could be measured, it could be mitigated by various interventions.
Possible measurements of distraction include eye ﬁxations off the road (measured by
eye-tracking devices), high trafﬁc density (measured by out-the-window video image
analysis), control actions (measured by spectral analysis of steering wheel adjustments,
movement of the foot off the accelerator pedal, and use of turn signals), weather
(measured by precipitation sensors, road surface sensors, and use of windshield
wipers), and excessive speed. Interventions could be warning signals (e.g., auditory
warnings in the cell phone causing the other party as well as the driver to be aware of
the trafﬁc density or distraction source), disabling of certain functions (cutting off the
cell phone, navigation system, radio, or whatever), and modiﬁcations to other systems
(making the intelligent cruise control follow at a greater distance from the lead vehi-
cle, or providing a forward collision warning sooner than would otherwise occur).
The National Highway Trafﬁc Safety Administration has sponsored several national
forums on the topic of driver distraction (Llaneras, 2000). The detailed research ques-
tions relating to distraction mitigation by automation are too numerous to list here.
Adaptive Automation in Air Trafﬁc Control
Unlike the aircraft ﬂight deck, the air trafﬁc control (ATC) center typically has not
been highly automated. Because of increased air trafﬁc, however, and given the need
for greater efﬁciency, there have been several proposals to modernize the ATC system
with increased automation. A National Research Council committee examined this
Human-Automation Interaction 111
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 111
issue of ATC modernization and made a number of recommendations for enhancing
the current system with judicious use of automation of certain controller functions
(Wickens, Mavor, Parasuraman, & McGee, 1998). Some of these recommendations
appear to have been followed by the FAA, which has begun the implementation of a
number of new automated systems designed to aid controllers in the safe and efﬁcient
separation of air trafﬁc.
One example of new ATC automation is the Center TRACON Automation System
(CTAS), which consists of a suite of software tools for aiding controllers. Hilburn,
Jorna, Byrne, and Parasuraman (1997) used one of these tools to examine the effects
of adaptive automation on the performance of air trafﬁc controllers. Previous research
found that automation that is introduced in an attempt to reduce the operator’s men-
tal workload often has the opposite effect: increasing mental workload during high
task demand periods, reducing it unnecessarily during periods of low demand,or both
(E. L. Wiener & Curry, 1980). This phenomenon, one of the “ironies of automation”
ﬁrst referred to by Bainbridge (1983) and which E. L. Wiener (1988) called “clumsy”
automation,could be mitigated if the onset of automation were linked to task demand.
Hilburn et al. tested experienced controllers on a high-ﬁdelity simulation of the
Brussels-Maastricht ATC sector. They varied the trafﬁc load and complexity to create
low and high task demand periods during a simulated work shift. The controllers were
provided with a decision aid for determining the optimal descent trajectories of air-
craft at the start of initial approach—the Descent Advisor (DA) of CTAS. The DA also
gave information on potential aircraft-to-aircraft or aircraft-airspace conﬂicts and
offered possible resolutions to the conﬂict.
Hilburn et al. used two conditions: in the static automation condition, the DA was
provided throughout the work period, irrespective of trafﬁc load, whereas in the adap-
tive automation condition, the DA was provided only at high trafﬁc load and not dur-
ing periods of low trafﬁc load. These researchers found signiﬁcant beneﬁts for con-
troller workload (as assessed using pupilometric and heart rate variability measures)
when the DA was provided adaptively during high trafﬁc loads, compared with when
it was available throughout (static automation) or only at low trafﬁc loads. In addi-
tion, the adaptive condition was associated with signiﬁcantly better system perform-
ance, as indicated by a lower difference between estimated and actual time of arrival
at the airport and higher throughput.
Interfaces for Adaptable Automation: Delegation
Although there is a fairly large body of empirical evidence pointing to the system
performance beneﬁts of adaptive automation (Inagaki, 2003; Parasuraman, 2000;
Scerbo, 2001), it is far from clear how well fully adaptive systems will perform in
practice.In adaptive systems, the decision to invoke automation or to return an auto-
mated task to the human operator is made by the system in real time using any of
the previously described methods. This immediately raises the issue of user accept-
ance of such a system. Human operators may be unwilling to accede to the “author-
ity” of a computer system that mandates when and what type of automation is or
is not to be used.
112 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 112
Also of concern is the potential for system unpredictability and its consequences
for operator performance. Billings and Woods (1994), for example, warned that truly
adaptive systems may be problematic because their behavior may not be predictable
to the user. To the extent that automation can hinder the operator’s situation aware-
ness by taking him or her out of the loop, unpredictably invoked automation by an
adaptive system may further impair the user’s understanding of the situation. How-
ever, if the automation were explicitly invoked or changed in mode by the user, then
presumably the unpredictability would be lessened.
But involving the human operator in making decisions about when and what to
automate can increase mental workload. Further, in a team situation, one team mem-
ber may reconﬁgure the system and the other team member(s) not know about it
(Moray, 1992). Or, if responsibility is not clearly assigned, one team member may mis-
takenly assume that another team member will take some necessary action.
Thus, there is a trade-off between increased unpredictability and increased work-
load in systems in which automation is invoked by the system or by the user, respec-
tively. Opperman (1994) characterized these alternatives as adaptive and adaptable
approaches to system design (see also Scerbo, 2001). The combined human-machine
system adapts to various contexts in both cases. However, in adaptive systems,automa-
tion determines and executes the necessary adaptations, whereas in adaptable systems,
the operator is in charge of the desired adaptations. The distinction is primarily one
In an adaptable system, the human always retains the authority to invoke or change
the automation, whereas in an adaptive system this authority is shared. In adaptable
systems, therefore, the human operator is like the supervisor of a human team who
delegates tasks to team members—or, in this case, to automation. The challenge for
developing delegation interfaces to a system is that the operator should be able to make
decisions regarding the use of automation in a way that does not create such high
workload that any potential beneﬁts of delegation are lost.
Delegation interfaces may allow adaptable automation to be implemented at a ﬂex-
ible and appropriate balance point in this trade-off space (C. A. Miller & Parasura-
man, in press).Humans should be able to delegate tasks to automation at times of their
own choosing and receive feedback on their performance. Delegation in this sense is
similar to that which occurs in successful human teams—for example, self-organizing
of roles on aircraft carriers (Rochlin, LaPorte, & Roberts, 1987; LaPorte, 1996). It rep-
resents a real-time approach to supervisory control (Sheridan, 1976, 1992b). The
human operator sets an objective, provides instructions (at a greater or lesser level of
detail), and then delegates or authorizes the automation to determine the best method
by which to proceed within the instructions toward the goal. Delegation should pro-
vide a highly ﬂexible method for the human supervisor to declare goals and provide
instructions and thereby choose how much or how little autonomy to impart to
automation on a moment-by-moment basis.
An example of such a delegation interface is the Playbook
—so named because it
is based on the metaphor of a sports team’s book of approved plays and the selection
from among those plays by the team leader—for example, the quarterback in Ameri-
can football—and their execution by the team members—that is, the other players
Human-Automation Interaction 113
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 113
(C. A. Miller & Parasuraman, in press; C. A. Miller, Pelican, & Goldman, 2000). The
Playbook interface facilitates the teaching of automation, an idea ﬁrst proposed by
Sheridan (1976, 1992a) when he suggested the supervisory control concept. This is
both a shared knowledge structure of tasks and their relationships within which task
performance can be discussed by human and automation and a language for the com-
munication of instructions.
The Playbook uses a hierarchical task model to provide a common language with
which a human supervisor may communicate goals and intents, and a Hierarchical
Task Network planning system (Erol, Hendler, & Nau, 1994) to understand, reason
over, and either critique or complete partial plans provided by the human. This form
of interface permits the operator to delegate tasks to automation at a wide variety of
functional levels of abstraction by provision of goals and of full or partial plans.
Finally, the Playbook streamlines the process of delegation by the human operator by
providing a compiled set of plans, or “plays,” with short, easily commanded labels that
can be further modiﬁed as needed. This is the critical aspect of the concept that allows
this form of adaptable automation not to increase the workload associated with dele-
gation, much as a sports team has an approved set of plays that facilitate task delega-
tion by the team leader.
There are two sources of evidence concerning the efﬁcacy of the delegation
approach to adaptable automation. First, a Playbook prototype for a mission planning
tool for commanding unmanned combat air vehicles (UCAVs) has been developed as
a proof of concept (C. A. Miller, Goldman, Funk, Wu, & Pate, 2004). Second, initial
experimental studies of the effects of delegation interfaces on human performance
have been carried out (Parasuraman, Galster, Squire, Furukawa, & Miller, 2005). These
studies examined the use of a simple delegation interface on system performance dur-
ing simulated human-robot teaming using the RoboFlag simulation environment.
RoboFlag provided the operator the ability to command simulated robots, individu-
ally or in groups,at several levels of detail: by providing designated endpoints for robot
travel, by commanding higher-level behaviors (or modes or plays) such as “Patrol Bor-
der” or “Circle Defense,” or by even higher “super-plays” such as “Go on Offense.” The
results showed that the multilevel tasking provided by the delegation interface allowed
effective user supervision of robots, as evidenced by the number of missions success-
fully completed and the time for mission execution. However, additional studies are
needed in which more complex versions of delegation interfaces are evaluated.
CURRENT AND FUTURE AUTOMATION
Future automation technology will be characterized by being smaller and smarter.
Miniaturization of sensors, actuators, and computers (nanotechnology) goes hand in
hand with lower power consumption. This will mean that the human user will com-
mand, directly or indirectly, many more automatic control loops than today. However,
most of these will not be apparent to the human; they will be hidden from view,much
as are the 20–40 computer chips in today’s automobiles. More and more of the analy-
114 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 114
sis and decision making will be provided by so-called computerized intelligent agents
(Weiss, 1999), but there are serious challenges in getting humans and automation to
cooperate, as will be discussed next.
Challenges in Various Application Sectors
Sectors of the economy where robot automation has been prevalent will see a contin-
uing trend toward more, smaller, and more capable robots in manufacturing plants,
in hostile environments such as undersea and in space, on farms, in homes, and in
A visitor to a modern automobile ﬁnal assembly plant will ﬁnd numerous robots
ﬁtting parts together (including glass and cloth or rubber components that only a few
years ago required skilled human labor) as well as welding and painting. Modern chip-
making and electronics assembly are done largely by robots, which have the advantage
of being faster, cleaner, and more precise than humans. In chemical processing and
genetic engineering, robotic manipulators combine with automatic measurement and
ﬂow control to perform batch processing, often on a scale (of complexity, not size)
that is far greater than what a human operator can handle, or far more delicate, and
in most cases much faster. Even in China, where labor is cheap by Western standards,
engineers are installing factory automation on a large scale.In all such cases the human
operator is relegated to the role of supervisor, as described previously.
The undersea robot (controlled remotely by a human) is gradually replacing the
human diver because it is able to go to the greatest ocean depths and to work con-
tinuously. Robots are now performing oil, gas, and mineral exploration and mining
operations in the ocean, where operations must gradually go deeper as land-based and
shallow-ocean reserves run out. Space robots have repeatedly proven themselves well
suited to planetary exploration and more recently have been favored over human astro-
nauts for performing delicate repairs on the Hubble space telescope and other valu-
able scientiﬁc instruments.
Most undersea and space robots do not perform repetitive tasks as do factory
robots; each new task is different. They are more appropriately called teleoperators (if
they are remotely controlled by humans continuously) or telerobots (if they are truly
autonomous at least for short periods and generally under supervisory control by their
remote human operators; see Sheridan, 1992a).
As world populations increase, deforestation continues, and arable lands slowly
become desert or are ﬂooded by rising ocean levels, the demand for food will increase.
A demand for more automation is bound to occur to make farming more efﬁcient.
Robotic tractors and harvesters are gradually replacing manned vehicles on farms.
Agricultural robotics will include robots to farm the ocean,which accounts for roughly
95% of the earth’s biosphere (by volume) and is a good source of both animal and
plant protein. However, despite all the progress in robotic sensing, mobility, and
manipulation,the robot brains are still not capable of being independent of the human
For the most part there has been no need to make robots look like people (anthro-
pomorphic, or humanoid), except for entertainment. However, roboticists have shown
Human-Automation Interaction 115
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 115
increasing interest in research to provide robots with mobility, gestures, and facial
expressions that resemble those of humans. This is in part a means to study human
behavior (the idea being that imitation begets understanding) and partly to improve
communication between robots and people; presumably a person can work better with
a machine that acts like a person, though this premise needs to be more fully explored
(Lewandowsky et al., 2000). Such “social robots” are also being designed for use with
special populations of people, such as the frail elderly, the disabled, or autistic persons.
Until now the World Wide Web has been used primarily for communicating text,
graphics, and software. A continuing application has been information searches, such
as are provided by Google and other search engines. These are becoming ever more
powerful for eliciting from a human what he or she wishes to ﬁnd and for presenting
that information in a manner tailored to the individual user. New applications of the
Web are to control machines from a distance—for example, to start food cooking on
a stove, to start the robotic lawn mower or vacuum cleaner, to control a robot to do
surveillance at a home or company or government property, or to remotely manipu-
late or position some object for examination through a television system.
One particular automation project of special interest to human factors engineers is
the effort to reduce the crews of ships, both navy and merchant marine. Traditionally,
personnel on ships have been specialized: Some tended engines, some scrubbed decks,
some managed cargo,and some cooked. On navy ships there were specialists for shoot-
ing guns and ﬁghting ﬁres. Ships of the future will require that crew members serve
multiple functions and, more than anything, that they be supervisors of automation.
Control will be mostly centralized in one command center. The conﬁguration of the
navy ship will be different, with very little space on open deck (practically nothing for
a human to do there!). Almost all ships can now be controlled by computer; GPS sen-
sors measure the ship’s position relative to established navigation databases of the har-
bors, including channels and hazards.
Cargo ships have already undergone much automation, and this trend will con-
tinue. Probably the most automated port is that at Rotterdam, where shipping con-
tainers are slid from conventional trucks onto a small number of driverless trucks,
which in turn deliver them under computer control to speciﬁed points adjacent to the
docked ships. Here huge robot arms pick up the containers and stack them on the
appropriate ship. Unloading and stacking for subsequent dispatch is similarly auto-
mated. Human operators perform their supervisory functions from a control tower
similar to that at an airport.
On naval ships there is also a general trend for fewer personnel to be available,
owing to both the all-volunteer force and demographic trends. Furthermore, the mul-
tiplicity of U.S. military commitments around the world means that personnel are
spread thin, and automated systems will increasingly take up the slack.
High-frequency radio technology has made many advances recently not only for cell
phones and wireless computer communication but also for many other applications.
The personal digital assistant (PDA) combining radio, MP3 player, cell phone,
pager, date book, and Internet access is gradually emerging in an easily wearable unit,
and in the future we may no longer need these functions embodied in separate items
in our homes. The ﬁrst experiments are now under way with highway vehicles that
116 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 116
communicate with one another and issue mutual warnings to vehicles converging at
intersections (which can also apply the brakes should the human drivers not be pay-
Another example of innovative automation is radio frequency identiﬁcation (RFID)
chips that, when energized from a few meters away, will communicate back a radio
signal. These are used on library books and retail merchandise to keep track of items
going in and out of libraries and stores. RFID chips are getting tinier and cheaper
(postage stamp size, and a penny or so in cost). Thus they can be pasted onto any
package to maintain inventory control from the factory to the customer, with each
RFID package uniquely identiﬁed by a string of bits. One can imagine all sorts of
future uses to enable people to keep track of items in the home—and children. Par-
ents in Korea are already keeping track of their children through graphic maps on their
cell phones, driven by the GPS chips in their children’s phones.
Automation is present in the hospital in many forms. Medical records are now com-
puterized in many hospitals and in the future will follow the patient between the physi-
cian’s ofﬁce and the hospital. They will even be available to the patient, with pointers
to explanatory material to educate the patient as to ﬁndings and interpretations. But
this communication also poses huge problems of privacy, similar to those that have
already occurred in ﬁnancial transactions.
Much more automated medical testing technology will soon ﬁnd its way into physi-
cians’ofﬁces and nursing homes, and even into the home for use by patients on them-
selves or by family members of home health care workers. It will need to be fail-safe
and easy for unsophisticated persons to use. Much-improved telecommunication
between homes and hospitals will occur. This will enable home-rendered tests to be
analyzed in hospitals and emergencies to be triaged so that patients will know when
there is an emergency and immediate hospitalization is required and when there is no
immediate cause for concern (Sheridan & Thompson, 1994).
Miniaturization also means that health sensors and alerting devices will increasingly
be worn and integrated with clothing or be implanted within the body. This form of
automation will have a growing market among older adults, who are making up an
increasing share of our population.
Homeland security has seen much new automation to detect explosives and
radioactive agents in baggage and on persons boarding aircraft. Similar new technol-
ogy may soon be used on trains and to control entry to public buildings, sporting
events, and so on. Improved personal identiﬁcation in some form will probably be in
wide use within the decade. As with health care technology, sensors to monitor per-
sonal security and alert family or police will ﬁnd new markets.
Challenges of Human Operator
Coordination with Automation
The new forms of automation that will pervade all aspects of life will pose new chal-
lenges for coordination between humans and automation. Woods (1996) pointed out
that adding automation with the intention of assisting the human operator is like
adding another team member, one who does not necessarily speak the same language
Human-Automation Interaction 117
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 117
and share the same cultural assumptions. This can result in what Woods has called
automation surprises (Sarter, Woods, & Billlings, 1997) and lead the human to ask E.
L. Wiener’s (1982) familiar questions: What is the automation doing now? Why is it
doing that? What will it do next?
Woods implies that increasing automation necessarily increases the demand for
coordination. It surely does so to a greater extent than most automation designers
appreciate. Yet greater coordination need not mean that the human cognitive load
increases over what is required for the no-automation case. If the automation can per-
form assigned tasks sufﬁciently reliably and transparently, and if the human operator
is sufﬁciently well trained so as to be easily able to observe what the automation is
doing, to understand why it is doing that, and to predict what it will do next, the cog-
nitive load can be diminished.
So the problem, as emphasized by Christoffersen and Woods (2002), is not one of
authority or autonomy but, rather, one of cooperation and observability. And coop-
eration between human and machine, just as between two humans, means shared rep-
resentations—where the operator’s mental representation (mental model) truly corre-
sponds to the functional and causal behavior of the machine and both correspond to
the physical representations of the operator’s interface (displays and controls).
Sheridan and Verplank (1978) discussed the importance in well-deﬁned tasks—for
example, in controlling a telerobot—of displays and hand controls that maintain kine-
matic correspondence to the posture and movement of the robot’s end effector using
so-called resolved motion algorithms, and thus abiding by the all-important design
principles of observability and stimulus-response compatibility. Christoffersen and
Woods noted that in older hardwired control rooms, multiple operators can infer what
others are doing by direct visual observation, often without having to resort to sym-
bolic voice or computer-mediated communication. Exactly how they do this, and how
any team of operators uses both body and speech communication to perform joint
cognitive activity—what Hutchins (1995) calls “cognition in the wild”—remains a hard
problem for anthropologists and psychologists to model in any objective way. If we
understood, maybe we could even endow our computer agents with a modicum of
Klein, Woods, Bradshaw, Hoffman, and Feltovich (2004) ampliﬁed these ideas by
posing 10 challenges for making automation a team player in joint human-agent activ-
ity. Their point of view is a “basic compact”or tacit agreement among the joint human
workers. The automation must be designed to buy into this compact, a design that
includes (a) common grounding, (b) the ability to model each others’ intents and
actions,(c) interpredictability,(d) amenability to direction,(e) an effort to make inten-
tions obvious, (f) observability, (g) goal negotiation, (h) planning and autonomy sup-
port, (i) attention management, and (j) cost control. As spelled out by the authors,
many of these desirable goals have some overlap and have long been discussed as short-
comings of automation and questioned as being within the state of the automation
Having a computer share assumptions with a human is not easy, given that every
human has a lifetime of cultural assumptions that can at times (especially in decision
trade-offs) bear on the task at hand. Humans have a fuzzy knowledge and rule base
118 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 118
that is mostly beyond the capability of software engineers to encode. Having a com-
puter model for predicting human intentions and behavior is a tall order. But these
are nevertheless worthy research and engineering challenges for the long term.
SOCIAL, POLITICAL, AND ETHICAL ISSUES
Reliability and Liability
When automation fails, there are serious issues about who is liable: the operator, the
ﬁrm or agency that owns the equipment and/or employs the operator, the designer,
the manufacturer, the installer, the maintainer, and so on.There is a tendency to blame
automation mishaps on the operator because he or she is typically present when an
accident occurs. A broader view suggests that operators should have some relief from
responsibility if the machine fails. Probably the wisest policy is to abstain from blam-
ing anyone, at least until a thorough investigation has taken place. Typically it is dis-
covered that multiple factors came together to “cause” an accident, and either the
blame should be shared or no one should be blamed: the mishap should be regarded
generally as a learning experience for all (Reason, 1990).
Often what the operator thinks was commanded of the automation (in the oper-
ating mode the operator thinks is operative) is not what actually was commanded. The
purported father of cybernetics, Norbert Wiener, in his Pulitzer Prize-winning book
(1964), asserted that as computers and automation become more complex and peo-
ple become more dependent on such systems for transportation, communication,
health care, and national security, there is a growing danger that the expectations of
the humans will not match the logic of the machine, where the latter dictates what the
automation will eventually do. Thus a degree of skepticism is critical when proponents
urge that large-scale systems such as air trafﬁc control and guided missile weapons sys-
tems be highly automated.
Virtual Reality and What Is Real
Computer graphics and display technology have made remarkable progress in recent
years, enabling simulation and accompanying immersion that gives the user the feel-
ing of actually being there. It makes human-in-the-loop simulators for aircraft, high-
way vehicles, ships,and minimally invasive surgery more realistic and therefore is more
acceptable than it once was for training and research. Researchers have debated how
and even whether such immersion is important for training (Darken & Goerger, 1999)
and whether the cost is justiﬁed (though it is obvious that it is important for enter-
tainment and seems to sell simulators). The computer game and movie/TV special
effects industry has motivated not only new software techniques that take advantage
of higher-speed computers and higher display resolutions but also new head-mounted
displays,“data gloves,” and 3-D auditory display techniques that enable the immersion
effect. The effect is enhanced not only by higher-resolution displays but also by
the user’s being able to change the viewpoint, head orientation, and hand position
Human-Automation Interaction 119
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 119
relative to a touched object and to have the sensory pattern change according to expec-
tation (Sheridan, 1992a).
In due course the technology will meet the criterion of the Turing test, whereby the
observer will not be able to discriminate what is virtual from what is real. That poses
serious concerns—for example, worries that children who spend hours playing violent
and very realistic computer games will transfer that violent behavior into their real-
world lives. Even at the seemingly more innocuous level of mechanized toys there is
concern that the toys’ use is too preprogrammed and children’s development of imag-
ination is inhibited (Sheridan, 1993).
The compelling nature of virtual reality is illustrated from a piece in Wired Today.
After a recent three-day binge of playing the Japanese cult hit video game Katamari
Damacy, Los Angeles artist Kozy Kitchens discovered that walking away from the game
was not as easy as putting down her joystick. In the game, players push around what
amounts to a giant tape ball, attempting to make the ball bigger by picking up any
and all objects in its path. Kitchens found that her urge to keep picking things up was
not so easy to shake.
“I was driving down Venice Boulevard,” recalled her husband, Dan Kitchens, “and
Kozy reached over and grabbed the steering wheel and for a moment was trying to
yank it to the right....(Then) she let go, but kept staring out her window, and then
looked back at me kind of stunned and said, ‘Sorry. I thought we could pick up that
mailbox we just passed.’”
Though motorists and pedestrians shouldn’t worry too much about rogue Kata-
mari Damacy players, Kozy Kitchens’ difﬁculty with separating her real-life con-
sciousness from that of her game playing is all too common among hard-core gamers.
It’s so common, in fact, that game publishers might want to consider warning their
customers that they may soon be unable to tell the difference between the game and
reality. Frequent gamer Alfred Weisberg-Roberts said he often feels lingering effects
after playing games like Animal Crossing, in which the point is to collect as many ani-
mals and bugs as possible from a wide variety of locations.
“Once, my girlfriend happened upon a tree . . . kind of like the round, thin trees
in the game, and began to shake it—one in-game way of receiving money, goods, and
bees,”Weisberg-Roberts said.“When nothing fell from its branches, I think she quickly
realized how this must have looked to the other hundred or so people in the park.”
Mixed-Initiative Conﬂicts within Large-Scale Systems
Large-scale systems are characterized by many sensors, many different computers as
well as people performing analyses and making decisions, and many actuators taking
physical actions to implement those decisions. This usually means that there are mul-
tiple control loops, each trying to drive some variable to correspond with its reference
input in spite of external disturbances. The problem is that within the physical process
being controlled, these actions can be and often are coupled—meaning the action of
one control loop appears as a disturbance to the other control loop. For example, in
a robotic system one control loop may be programmed to drive the robot to move
toward a target, and a second control loop is programmed to avoid obstacles. If the
120 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 120
robot runs into an obstacle while moving toward the target, there is an impasse. Fig-
ure 2.3 illustrates the idea; the heavy arrows designate the two automated control
loops. This is commonly called a mixed-initiative problem.
Human supervisors may have access to the same sensed information as the automa-
tion they are overseeing but not to information from the other control loop. In other
words, they see different sides of the elephant. If the two supervisors had access to
each other’s information, in theory they might be able to resolve the conﬂict, but the
additional information just adds to their workload, and furthermore the communica-
tion between them (dashed line) is likely to be delayed, noisy, or nonexistent.
Another type of conﬂict arises when the actuators are automatically drawing from
a common resource pool (or supplies, energy, budgeted money, etc.) and each super-
visor is trying to fulﬁll his or her own goals independent of the others. This is what
Hardin (1968) has called “the tragedy of the commons.”Naturally, when there are more
control loops and more people in the system, these problems are compounded.
Automation as a Cause of User Alienation
We use the term alienation to characterize a number of social and political issues con-
cerning the individual. In an automated system the human is removed not only spa-
tially but also temporally, functionally, and cognitively from the ongoing physical
process he or she is directing. What the human does and thinks is likely to have dif-
ferent timing, different explicit form, and different logical content from what the
Human-Automation Interaction 121
Figure 2.3. Mixed-initiative conﬂicts within large-scale automation.
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 121
automation does and thinks. Various authors have discussed the alienation problem
(e.g., see Hofstede, 1994; Moray, 2000; Zuboff, 1988). Sheridan (1980) presented some
components of alienation, as follows.
Threatened or actual unemployment. Organizations have become more efﬁcient
in their use of human supervisors: Fewer people can now supervise more machines.
Furthermore,automation is becoming more able to detect its own failures,and in some
cases it can even repair itself. The real threat has been unemployment of the unskilled
and technologically illiterate. However, with the Internet breaking down geographic
barriers, even the skilled and technologically literate are in danger of unemployment
as high-skill jobs move to technologically sophisticated regions in Asia. The threat of
technological displacement has moved up the chain, from the manual laborer and rail-
road brakeman to the computer system administrator, radiologist, and architect.
Erratic mental workload and dissatisfaction with work. Automation affects not
only the nature but also the pace of work and may at times make that pace vary
between extremes. This is the oft-cited “hours of boredom punctuated by moments of
Centralization of management control and loss of worker control. A result of
automation and associated electronic technology is that management can secretly
record and monitor workers. The mere possibility of being monitored in this way is
often sufﬁcient to produce worker anxiety, including fear that private data stored
electronically may be accessed by persons other than those authorized. In some cases
centralized monitoring may enhance productivity, but in other cases it may prove
Desocialization. Interaction with computers is gradually replacing interaction with
other people. As supervisory control systems are interconnected, the computer will
mediate increasingly more of what interpersonal contact remains, as has already hap-
pened in many cases with e-mail, pagers, and associated software for management
coordination and computer-supported cooperative work.
Deskilling. Skilled workers who are “promoted” to supervisory controllers (some-
times derogatorily referred to as “button pushers”) may resent the transition. In part,
this may be out of fear that when called on to take over and do the job manually, they
may not be able to do so.
Intimidation of greater power. Automation encourages larger aggregations of
interconnected equipment, higher speeds, greater complexity, and probably greater
economic risk if something goes wrong and the supervisor doesn’t take the appropri-
ate corrective action. The human supervisor will be forced to assume increasingly more
ultimate responsibility, although in most cases the responsibility probably should
reside with some combination of the manager and the system designer.
122 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 122
Technological illiteracy. In the role of supervisory controller, the operator may lack
technological understanding of how the computer and the rest of the complex tech-
nology do what they do. What is really going on with the communications and con-
trol software may be too specialized even for many technicians involved with the newer
systems. The push toward increasing functionality and capability may mean that sys-
tems are becoming increasingly less visible and comprehensible to even the designers
Mystiﬁcation and misplaced trust. Human operators of computer-based systems
sometimes become mystiﬁed by and superstitious about the power of the computer,
even seeing it as a kind of magic or a “big brother” authority ﬁgure. This leads natu-
rally to misplaced trust.
Sense of not contributing. Though the efﬁciency and mechanical productivity of a
new supervisory control system may far exceed that of an earlier manually controlled
system, the operator may come to feel that with automation he or she is no longer the
source of value added, no longer a signiﬁcant contributor. The sense of personal pro-
ductivity—what psychologist Erich Fromm (1995) called the productive orientation—
is allegedly fundamental to humans’ sense of self-worth. Without it, who are we?
Abandonment of responsibility. As a result of the factors just described, human
supervisors of automation may eventually feel they are no longer responsible for what
happens but that the computers are. A worker with his or her own set of hand tools
or a simple, self-powered, but manually controlled machine—though he or she may
sometimes place the blame for difﬁculties elsewhere—has a clear responsibility for use
and maintenance of the tools or machine. When workers’ actions are mediated by a
powerful computer, however, the lines of responsibility are not so clear, and the work-
ers may, in effect,abandon their responsibility for the task performed or the good pro-
duced, believing instead that it is in the “hands” of the computer.
Blissful enslavement. To many writers the worst form of alienation, the worst
tragedy, occurs when a worker is happy to accept a role in which he or she is made to
feel powerful but, in actuality, he or she is enslaved. Both Aldous Huxley’s Brave New
World and George Orwell’s 1984 are famous for this theme of blissful enslavement.
Ease of Committing Violent Acts by
Remote Control of Automation
Today’s automation coupled with modern communication technology means that any-
one can supervise an automatic machine from any arbitrary distance away—and with-
out the source of control being apparent. This has been a boon to developers of
unmanned space, aerial, undersea, and land vehicles, most of which have been devel-
oped for military use. In contrast to past warfare, in which the ﬁghter risked his or
her own life for the cause, today remote-control capabilities mean that violence can
Human-Automation Interaction 123
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 123
be committed anonymously by anyone without the perpetrator’s even being aware of
the result—for example, many innocents may be killed because a highly automated
and even precisely targeted missile encountered circumstances on the ground that were
unpredicted. No one likes terrorists with car bombs or bombs strapped to their bod-
ies. But we can look forward to a day when violent acts or just insidious spying can
be committed by unmanned vehicles and robots controlled by anyone able to acquire
the communication and control technology, which will surely be getting smaller and
cheaper. Social and moral responsibility increases as automation affects more people
in more profound ways.
So What Should Not Be Automated?
At the end of this chapter we directly pose the question that is often asked of profes-
sionals concerned with humans and automation: What should not be automated, even
though it is possible? Automation engineers—at least many of them—have the atti-
tude that if it is technologically, economically, and legally feasible to automate, then
do it; it is an exciting challenge (Bainbridge, 1983). Much of the foregoing discussion
suggests reasons to go slow, to anticipate and evaluate the much more subtle problems
that automation can bring. The human factors engineer is often regarded by the engi-
neer as a worry-wart, a nay-sayer, a wet blanket in terms of risk taking and progress.
But the assessment of human-related questions—potential effects on individual and
social behavior, institutions, and culture—must be asked because, after all, the ultimate
purpose of technology is to make life better for people (Hancock, 1996). Furthermore,
these hard questions must be asked early—before the technological development is too
far along, before the point of no return, or at least before the point where changes are
very much more expensive than they would have been had they been made early.
There is a belief among many automation engineers that one can eliminate human
error by eliminating the human operator. To the extent a system is made less vulner-
able to operator error, it is made more vulnerable to designer error (Parasuraman &
Riley, 1997). And given that the designer is also human, this simply displaces the locus
of human error. In the end, automation is really human after all.
Abbott, K., Slotte, S., Stimson, D., Amalberti, R. R., Bollin, G., Fabre, F., et al. (1996). The interfaces between
ﬂightcrews and modern ﬂight deck systems (Report of the FAA Human Factors Team). Washington, DC:
Federal Aviation Administration.
Aviation coding manual. (1998). Retrieved October 10, 2005, from http://www.ntsb.gov/aviation/
Aviation Safety Reporting System. (2005). Retrieved October 10, 2005, from http://asrs.arc.nasa.gov/
Bagheri, N., & Jamieson, G. A. (2004). Considering subjective trust and monitoring behavior in assessing
automation-induced “complacency.” In D.A. Vicenzi, M. Mouloua, & P. A. Hancock (Eds.), Human per-
formance, situation awareness, and automation: Current research and trends (pp. 54–59). Mahwah, NJ:
Bainbridge, L. (1983). Ironies of automation. Automatica, 19, 775–779.
124 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 124
Barnes, M., & Grossman, J. (1985). The intelligent assistant concept for electronic warfare systems (Technical
Report NWC TP 5585). China Lake, CA: Naval Warfare Center.
Bennett, K.,& Flach, J. (1992). Graphical displays: Implications for divided attention, focused attention, and
problem solving. Human Factors, 34, 513–533.
Billings, C. E. (1997). Aviation automation: The search for a human centered approach. Mahwah,NJ: Erlbaum.
Billings, C. E., Lauber, J. K., Funkhouser, H., Lyman, G., & Huff, E. M. (1976). NASA aviation safety report-
ing system (Technical Report TM-X-3445). Moffett Field, CA: NASA Ames Research Center.
Billings, C. E., & Woods, D. D. (1994). Concerns about adaptive automation in aviation systems. In R. Para-
suraman & M. Mouloua (Eds.), Human performance in automated systems: Current research and trends.
(pp. 264–269). Mahwah, NJ: Erlbaum.
Byrne, E. A., & Parasuraman, R. (1996). Psychophysiology and adaptive automation. Biological Psychology,
Christoffersen, K., & Woods, D. D (2002). How to make automated system team players. In E. Salas (Ed.),
Advances in human performance and cognitive engineering research (vol. 2, pp. 1–12). Amsterdam: Elsevier.
Comstock, J. R., & Arnegard, R. J. (1992). Multi-attribute task battery (NASA Technical Memorandum
104174). Hampton, VA: NASA Langley Research Center.
Craik, K. J. W. (1947). Theory of the human operator in control systems, I: The operator as an engineering
system. British Journal of Psychology, 38, 56–61.
Darken,R. P., & Goerger, S. R.(1999). The transfer of strategies from virtual to real environments:An expla-
nation for performance differences? In Proceedings of Virtual Worlds and Simulation ’99 (pp. 159–164).
La Jolla, CA: Society for Computer Simulation International.
Degani, A. (2003). Taming Hal: Designing interfaces beyond 2001. New York: Palgrave MacMillan.
Dialogs on function allocation [Special issue]. (2000). International Journal of Human-Computer Studies (Vol
Endsley, M.,& Kaber, D. (1999).Level of automation effects on performance, situation awareness and work-
load in a dynamic control task. Ergonomics, 42, 462–492.
Erol, K., Hendler, J., & Nau, D. (1994). UMCP: A sound and complete procedure for hierarchical task net-
work planning. In K. Hammond (Ed.), AI planning systems: Proceedings of the 2nd International Confer-
ence (pp. 249–254). Los Altos, CA: AAAI.
Fitts, P. M. (2005). Some basic questions in designing an air-navigation and air-trafﬁc control system. In N.
Moray (Ed.), Ergonomics major writings (Vol. 4, pp. 367–383). London: Taylor & Francis. (Washington
DC, Reprinted from Human engineering for an effective air navigation and trafﬁc control system, National
Research Council, pp. 5–11, 1951, Washington, DC: National Academy Press.)
Fromm, E. (1995). Escape from freedom. New York: Holt.
Funk, K., Lyall, B., Wilson, J., Vint, R., Miemcyzyk, M., Suroteguh, C., et al. (1999). Flight deck automation
issues. International Journal of Aviation Psychology, 9, 125–138.
Furukawa,H., Parasuraman, R.,& Inagaki, T. (2003). Supporting system-centered view of operators through
ecological interface design: Two experiments on human-centered automation. In Proceedings of the
Human Factors and Ergonomics Society (pp. 567–571).Santa Monica,CA: Human Factors and Ergonom-
Gibson, J. J. (1979). The ecological approach to visual perception. Boston: Houghton-Mifﬂin.
Green, D. M., & Swets, J. A. (1966). Signal detection theory and psychophysics. New York: Wiley.
Grice, H. P. (1975). Logic and conversation. In P. Cole & J. Morgan (Eds.), Syntax and semantics: Speech acts
(Vol. 3, pp. 276–290). New York: Academic.
Hammer, J. M., & Small, R. L. (1995). An intelligent interface in an associate system. In W. B. Rouse (Ed.),
Human/technology interaction in complex systems (Vol. 7, pp. 1–44). Greenwich, CT: JAI Press.
Hancock, P. A. (1996). Teleology for technology. In R. Parasuraman & M. Mouloua (Eds.), Automation and
human performance: Theory and applications (pp. 461–497). Mahwah, NJ: Erlbaum.
Hancock, P. A., Chignell, M. H., & Lowenthal, A. (1985). An adaptive human-machine system. In Proceed-
ings of the IEEE Conference on Systems, Man and Cybernetics, 15 (pp. 627–629). Washington, DC: IEEE.
Hancock, P. A., & Scallen, S. F. (1996, October). The future of function allocation. Ergonomics in Design
Hardin, G. (1968). The tragedy of the commons. Science, 162, 1243–1248.
Human-Automation Interaction 125
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 125
Hick,W. E.(1952). On the rate of gain of information. Quarterly Journal of Experimental Psychology, 4, 1–26.
Hilburn, B., Jorna, P. G., Byrne, E. A., & Parasuraman, R. (1997). The effect of adaptive air trafﬁc control
(ATC) decision aiding on controller mental workload. In M. Mouloua & J. Koonce (Eds.), Human-
automation interaction (pp. 84-91). Mahwah, NJ: Erlbaum.
Hofstede, G. (1994). Cultures and organizations. London: HarperCollins.
Hutchins, E. (1995). Cognition in the wild. Cambridge: MIT Press.
Inagaki, T. (2003). Adaptive automation: Sharing and trading of control. In E. Hollnagel (Ed.), Handbook
of cognitive task design (pp. 221–245). Mahwah, NJ: Erlbaum.
James, H., Nichols, N., & Phillips, R. (1947). Manual tracking. In Theory of servomechanisms. New York:
Jamieson, G. A., & Vicente, K. J. (2005). Designing effective human-automation-plant interfaces: A control
theoretic perspective. Human Factors, 47, 12–34.
Jordan, N. (1963). Allocation of functions between man and machines in automated systems. Journal of
Applied Psychology, 47, 161–165.
Kaber, D. B., & Endsley, M. (2004). The effects of level of automation and adaptive automation on human
performance, situation awareness and workload in a dynamic control task. Theoretical Issues in Ergonom-
ics Science, 5, 113–153.
Kaber, D. B., & Riley, J.M. (1999). Adaptive automation of a dynamic control task based on workload assess-
ment through a secondary monitoring task. In M. Scerbo & M. Mouloua (Eds.), Automation technology
and human performance: Current research and trends (pp. 129–133). Mahwah, NJ: Erlbaum.
Klein, G., Woods, D. D., Bradshaw, J. M., Hoffman, R. R., & Feltovich, P. J. (2004). Ten challenges for mak-
ing automation a team player in joint human-agent activity. IEEE Intelligent Systems 19(6), 91–95.
Kleinman, D. L., Baron, S., & Levison, W. H. (1970). An optimal control model of human response, Part 1.
Automatica, 63, 357–369.
LaPorte, T. R. (1996). High reliability organizations: Unlikely, demanding, and at risk. Journal of Contin-
gencies and Crisis Management, 4(2), 60–71.
Lee, J. D., & Moray, N. (1992). Trust, control strategies and allocation of function in human-machine sys-
tems. Ergonomics, 35, 243–1270.
Lee, J., & Moray, N. (1994). Trust, self conﬁdence, and operators’ adaptation to automation. International
Journal of Human-Computer Studies, 40, 153–184.
Lee, J., & See, J. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46, 50–80.
Lee, J. D., Caven, D., Haake, S., & Brown, T. L. (2001). Speech-based interaction with in-vehicle computers:
The effect of speech-based e-mail on drivers’ attention to the roadway. Human Factors, 43, 631–640.
Lee,J. D., McGehee,D., Brown, T. L.,& Reyes, M.(2002). Collision warning timing, driver distraction, and driver
response to imminent rear end collision in a high-ﬁdelity driving simulator. Human Factors, 44, 314–334.
Lewandowsky, S., Mundy, M., & Tan, G. P. (2000). The dynamics of trust: Comparing humans to automa-
tion. Journal of Experimental Psychology: Applied, 6, 104–123.
Llaneras, R. E. (2000). NHTSA driver distraction internet forum. Retrieved October 10, 2005, from
McRuer, D. T., & Jex, H. R. (1967). A review of quasi-linear pilot models. IEEE Transactions on Human Fac-
tors in Electronics, HFE-4(3), 231–249.
McRuer, D., & Krendel, E. (1959). The human operator as a servo system. Journal of the Franklin Institute,
Metzger, U., & Parasuraman, R. (2001). Automation-related “complacency”: Theory, empirical data, and
design implications. In Proceedings of the Human Factors and Ergonomics Society 45th Annual Meeting
(pp. 463–467). Santa Monica, CA: Human Factors and Ergonomics Society.
Metzger, U., & Parasuraman, R. (2005). Automation in future air trafﬁc management: Effects of decision aid
reliability on controller performance and mental workload. Human Factors, 47, 35–49.
Meyer, J. (2001). Effects of warning validity and proximity on responses to warnings. Human Factors, 43,
Miller, C. A. (2004). Human-computer etiquette [Special issue]. Communications of the ACM, 37(4).
Miller, C. A., Goldman, R., Funk, F.,Wu, P., & Pate, B. (2004, June). A Playbook approach to variable auton-
omy control: Application for control of multiple, heterogeneous unmanned air vehicles. In Proceedings
of Forum 60, the Annual Meeting of the American Helicopter Society. Alexandria, VA: AHS International.
126 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 126
Miller, C. A., & Hannen, M. D. (1999). The rotorcraft pilot’s associate: Design and evaluation of an intelli-
gent user interface for cockpit information management. Knowledge-Based Systems, 12, 443–456.
Miller, C. A., & Parasuraman, R. (in press). Designin for ﬂexible interaction between humans and automa-
tion. Human Factors.
Miller, C. A., Pelican, M., & Goldman, R. (2000). “Tasking” interfaces for ﬂexible interaction with automa-
tion: Keeping the operator in control. In Proceedings of the Conference on Human Interaction with Com-
plex Systems (pp. 123–128). Urbana-Champaign, IL: HICS.
Miller, G. A. (1956). The magical number 7, plus or minus 2: Some limits on our capacity for processing
information. Psychological Review, 63, 81–97.
Molloy, R.,& Parasuraman, R.(1994).Automation-induced monitoring inefﬁciency: The role of display inte-
gration and redundant color coding. In M. Mouloua & R. Parasuraman (Eds.), Human performance in
automated systems: Current research and trends (pp. 224–228). Mahwah, NJ: Erlbaum.
Moray, N.(1986). Monitoring behavior and supervisory control. In K. Boff, L. Kaufman, & J. Thomas (Eds.),
Handbook of perception and human performance (vol. 2, pp. 40-1–40-51). New York: Wiley.
Moray, N. (1992). Flexible interfaces can promote operator error. In J. Kragt (Ed.), Case studies in ergonom-
ics (pp. 49–64). London: Taylor & Francis.
Moray, N. (2000). Culture, politics and ergonomics. Ergonomics, 43, 858–868.
Moray, N. (2005). Ergonomics major writings. London: Taylor & Francis.
Moray, N.,& Inagaki,T. (2001).Attention and complacency. Theoretical Issues in Ergonomics Science, 1, 354–365.
Moray, N., Inagaki, T., & Itoh, M. (2000). Situation adaptive automation, trust and self-conﬁdence in fault
management of time-critical tasks. Journal of Experimental Psychology: Applied, 6(1), 44–58.
Morrison, J.G., & Gluckman, J. P. (1994). Deﬁnitions and prospective guidelines for the application of adap-
tive automation. In M. Mouloua & R. Parasuraman (Eds.), Human performance in automated systems:
Current research and trends (pp. 256–263). Mahwah, NJ: Erlbaum.
Mosier, K., Skitka, L. J., Heers, S., & Burdick, M. (1998). Automation bias: Decision making and perform-
ance in high-tech cockpits. International Journal of Aviation Psychology, 8, 47–63.
Muir, B. M. (1988). Trust between humans and machines, and the design of decision aids. In E. Hollnagel,
G. Mancini, & D. D. Woods (Eds.), Cognitive engineering in complex dynamic worlds (pp. 71–84). Lon-
Nass, C., Moon,Y., Fogg, B. J., Reeves, B., & Dryer, D. C. (1995). Can computer personalities be human per-
sonalities? International Journal of Human-Computer Studies, 43, 223–239.
National Transportation Safety Board. (1973). Eastern Air Lines, Inc., L-1011, N310EA, Miami, Florida,
December 29, 1972 (AAR-73-14). Washington, DC: Author.
National Transportation Safety Board. (1998a). Brief of accident NYC98FA020. Washington, DC: Author.
National Transportation Safety Board. (1998b). Safety recommendation letter A-98-3 through -5, January 21,
1998. Washington, DC: Author.
Norman, D. A. (1990). The problem with automation: Inappropriate feedback and interaction, not “over-
automation.” Philosophical Transactions of the Royal Society (London), B237, 585–593.
Opperman, R. (1994). Adaptive user support. Mahwah, NJ: Erlbaum.
Parasuraman, R. (1993). Effects of adaptive function allocation on human performance. In D. J. Garland &
J. A. Wise (Eds.), Human factors and advanced aviation technologies (pp. 147–157). Daytona Beach, FL:
Embry-Riddle Aeronautical University Press.
Parasuraman, R. (2000). Designing automation for human use: Empirical studies and quantitative models.
Ergonomics, 43, 931–951.
Parasuraman, R., Bahri, T., Deaton, J. E., Morrison, J. G., & Barnes, M. (1992). Theory and design of adap-
tive automation in aviation systems (Technical Report, Code 6021). Warminster, PA: Naval Air Develop-
Parasuraman, R., & Byrne, E. A. (2003). Automation and human performance in aviation. In P. Tsang & M.
Vidulich (Eds.), Principles of aviation psychology (pp. 311–356). Mahwah, NJ: Erlbaum.
Parasuraman, R., Galster, S., Squire, P., Furukawa, H., & Miller, C. A. (2005). A ﬂexible delegation interface
enhances system performance in human supervision of multiple autonomous robots: Empirical studies
with RoboFlag. IEEE Transactions on Systems, Man & Cybernetics 35, 481–493.
Parasuraman, R., & Miller, C. (2004). Trust and etiquette in high-criticality automated systems. Communi-
cations of the Association for Computing Machinery, 47(4), 51–55.
Human-Automation Interaction 127
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 127
Parasuraman, R.,Molloy, R., & Singh, I. L. (1993). Performance consequences of automation-induced “com-
placency.” International Journal of Aviation Psychology, 3, 1–23.
Parasuraman, R.,& Mouloua, M.(1996). Automation and human performance: Theory and applications. Mah-
wah, NJ: Erlbaum.
Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse and abuse. Human Fac-
tors, 39, 230–253.
Parasuraman, R., Sheridan, T., & Wickens, C. (2000). A model for types and levels of human interaction
with automation. IEEE Transactions on Systems, Man and Cybernetics, SMC-30(3), 286–297.
Prinzel, L. J., Freeman, F. G., Scerbo, M. W., Mikulka, P. J., & Pope, A. T. (2000). A closed-loop system for
examining psychophysiological measures for adaptive automation. International Journal of Aviation Psy-
chology, 10, 393–410.
Rasmussen, J. (1986). Information processing and human-machine interaction. Amsterdam: North-Holland.
Rasmussen, J., Pedersen, A.-M., & Goodstein, L. (1995). Cognitive engineering: Concepts and applications.
New York: Wiley.
Reason, J. T. (1990). Human error. Cambridge, England: Cambridge University Press.
Reeves, B., & Nass, C. (1996). The media equation: How people treat computers, television, and new media
like real people and places. New York: Cambridge University Press.
Riley, V. (1996). Operator reliance on automation: Theory and data. In R. Parasuraman & M. Mouloua
(Eds.), Automation and human performance: Theory and applications (pp. 19–35). Mahwah, NJ:
Rochlin, E., LaPorte, T., & Roberts, K. (1987, Autumn). The self-designing high reliability organization: Air-
craft ﬂight operation at sea. Naval War College Review, 76–91.
Rouse, W. B. (1976). Adaptive allocation of decision making responsibility between supervisor and com-
puter. In T. B. Sheridan & G. Johannsen (Eds.), Monitoring behavior and supervisory control (pp.
295–306). New York: Plenum.
Rouse, W. B. (1980). System engineering models of human-machine interaction. Amsterdam: North-Holland.
Rouse, W. B. (1988). Adaptive aiding for human/computer control. Human Factors, 30, 431–438.
Sarter, N., Woods, D., & Billings, C. E. (1997). Automation surprises. In G. Salvendy (Ed.), Handbook of
human factors and ergonomics (2nd ed., pp. 1926–1943). New York: Wiley.
Sarter, N.,& Woods, D. D. (1995). How in the world did we ever get into that mode? Mode error and aware-
ness in supervisory control. Human Factors, 37, 5–19.
Scerbo, M. W. (1996). Theoretical perspectives on adaptive automation. In R. Parasuraman & M. Mouloua
(Eds.), Automation and human performance: Theory and applications (pp. 37–63). Mahwah, NJ: Erlbaum.
Scerbo, M. W. (2001). Adaptive automation. In W. Karwowski (Ed.), International encyclopedia of ergonom-
ics and human factors (pp. 1077–1079). London: Taylor & Francis.
Scerbo, M., Freeman, F., Mikulka, P. J., Di Nocera, F., Parasuraman, R., & Prinzel, L. (2001). The efﬁcacy of
physiological measures for implementing adaptive technology (NASA Technical Memorandum). Hampton,
VA: NASA Langley Research Center.
Senders, J. W. (1964). The human operator as a monitor and controller of multidegree of freedom systems.
IEEE Transactions on Human Factors in Electronics, HFE-5, 1–6.
Shannon, C. E. (1947). Communication in the presence of noise. Proceedings of the IRE, 37, 10–22.
Sheridan, T. B. (1960). The human metacontroller. In Proceedings of the Annual Conference on Manual Con-
trol. Stamford, CT: Dunlap Associates.
Sheridan, T. B. (1976). Toward a general model of supervisory control. In T. B. Sheridan & G. Johannsen
(Eds.), Monitoring behavior and supervisory control (pp. 271–282). Elmsford, NY: Plenum.
Sheridan, T. B. (1980, October). Computer control and human alienation. Technology Review, 60–73.
Sheridan, T. (1988). Trustworthiness of command and control systems. In Proceedings of the International Fed-
eration of Automatic Control Symposium on Man-Machine Systems (pp.427–431). Elmsford, NY: Pergamon.
Sheridan, T. B. (1992a). Musings on telepresence and virtual presence. Presence: Teleoperators and Virtual
Environments, 1(1), 120–126.
Sheridan, T. B. (1992b). Telerobotics, automation, and human supervisory control. Cambridge: MIT Press.
Sheridan, T. B. (1993). My anxieties about virtual environments. Presence: Teleoperators and Virtual Envi-
ronments, 2(2), 141–142.
128 Reviews of Human Factors and Ergonomics
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 128
Sheridan, T. B. (1998). Allocating functions rationally between humans and machines. Ergonomics in Design,
Sheridan, T. B. (2000). Function allocation: Algorithm, alchemy, or apostasy? International Journal of
Human-Computer Studies, 52, 203–216.
Sheridan, T. B. (2002). Humans and automation: Systems design and research issues. Santa Monica/New York:
Human Factors and Ergonomics Society/Wiley.
Sheridan, T. B., & Ferrell, W. R. (1974). Man-machine systems. Cambridge: MIT Press.
Sheridan, T., & Parasuraman, R. (2000). Human vs. automation in responding to failures: An expected value
analysis. Human Factors, 42, 403–407.
Sheridan, T. B., & Thompson, J. M. (1994). People vs. computers in medicine. In S. Bogner (Ed.), Human
error in medicine. Mahwah, NJ: Erlbaum.
Sheridan, T. B., & Verplank, W. L. (1978). Human and computer control of undersea teleoperators (Man-
Machine Systems Lab Report). Cambridge: Massachusetts Institute of Technology.
Sherry, L., & Polson, P. G. (1999). Shared models of ﬂight management system vertical guidance. Interna-
tional Journal of Aviation Psychology, 9, 139–153.
St. John, M., Kobus, D. A., Morrison, J. G., & Schmorrow, D. (2004). Overview of the DARPA augmented
cognition technical integration experiment. International Journal of Human-Computer Interaction, 17,
Swets, J. (1996). Signal detection theory and ROC analysis in psychology and diagnostics. Mahwah, NJ: Erl-
Vicente, K. J. (2002). Ecological interface design: Progress and challenges. Human Factors, 44, 62–78.
Vicente, K., & Rasmussen, J. (1992). Ecological interface design: Theoretical foundations. IEEE Transactions
on Systems, Man and Cybernetics, SMC-22, 589–606.
Wei, Z., Macwan, A., & Wierenga, P. (1998). A quantitative measure for degree of automation, and its rela-
tion to system performance. Human Factors, 40, 277–295.
Weiss, G. (Ed.). (1999). Multi-agent systems. Cambridge: MIT Press.
Wickens, C.D.,& Hollands, J. G.(2000).Engineering psychology and human performance. Upper Saddle River,
NJ: Prentice Hall.
Wickens, C. D., Mavor, A., Parasuraman, R., & McGee, J. (1998). The future of air trafﬁc control: Human
operators and automation. Washington, DC: National Academy Press.
Wiegmann, D. A., & Shappell, S. A. (1997). Human factors analysis of post-accident data: Applying theo-
retical taxonomies of human error. International Journal of Aviation Psychology, 7, 67–81.
Wiener, E. L. (1982). Human factors of advanced technology: “Glass cockpit” transport aircraft (NASA Con-
tractor Report 177528). Moffett Field, CA: NASA Ames Research Center.
Wiener, E. L. (1988). Cockpit automation. In E. Wiener (Ed.), Human factors in aviation. San Diego, CA:
Wiener, E. L., & Curry, R. E. (1980). Flight-deck automation: Promises and problems. Ergonomics, 23,
Wiener, N. (1964). God and Golem, Inc. Cambridge: MIT Press.
Wilson, G. F., & Russell, C. A. (2003). Real-time assessment of mental workload using psychophysiological
measures and artiﬁcial neural networks. Human Factors, 45, 635–643.
Woods, D. D. (1996). Decomposing automation: Apparent simplicity, real complexity. In R. Parasuraman &
M. Mouloua (Eds.), Automation technology and human performance. Mahwah, NJ: Erlbaum.
Zuboff, S. (1988). In the age of the smart machine: The future of work and power. New York: Basic Books.
Human-Automation Interaction 129
02_hfes_089_129_ch02 6/24/06 12:40 AM Page 129