Motivation
Imagine an entitled nephew whom after many years of expecting to inherit his
wealthy uncle's fortune comes to learn that he has been cut out of his uncle's will.
Infuriated, he puts his ri e in his car, hops in, and races toward his uncle's house.
On the way he is consumed with thoughts of shooting his uncle and the resultant
bloody carnage. With a *thud*, he is broken out of his fantasy as his car careens
into an unnoticed pedestrian. Upon getting out of the car he realizes he has run
down and killed none other than his uncle! The question then is whether the
killing was intentional1. Let us re-run the thought experiment, but allow for the
fact that something salient in the environment re-centers the nephew on driving
such that the pedestrian is identi ed prior to impact and speci cally identi ed
as the nephew's uncle. At this point, the nephew decides to modify his intention
to shoot his uncle by merely running him over instead. Say he does so. Did the
nephew kill his uncle intentionally in this version of the scenario?
In the case of the angry nephew, intuition suggests that the rst variant is
unintentional killing while the second variant is textbook intentional killing. The
inference to intentionality hinges on the reader's knowledge that the nephew was
distracted in the rst variant and thus unaware of what he was doing while it was
? The authors would like to acknowledge generous support from the O ce of Naval
Research under Award N0001418WX00166. The views expressed in this paper are
solely the authors and should not be taken to re ect any o cial policy or position
of the United States Government or the Department of Defense.
1 This example is a variant on the original thought experiment presented in [
4
]
happening. Lack of awareness is one critical factor that can undermine agentive
control, or the ability of the agent to appropriately guide action.
2
Attention and Intentional Action
Rather than wade directly into the murky waters of consciousness, we have
chosen to start with attention, which has a close relationship to conscious awareness
[
5
]. To begin giving a computational account of intentional action we draw upon
work by the philosopher Wayne Wu, who situates attention as mediating the
relationship between having an intention and successful intentional action [
7
].
Doing so makes room for agentive control, such that an agent merely having the
appropriate beliefs, desires, and intentions isn't su cient in all cases for
successful intentional action. Rather, attention must be paid in appropriate measure
to the task at hand to intentionally succeed at achieving one's goals, as was
demonstrated in the case of the angry nephew.
Elaborating on work by Wu, we have developed a novel computational
treatment of intentions that treat them as being partially constituted by attentional
priorities in the ARCADIA computational cognitive system [
3
]. Each intention
consist of high-level semantic information describing its conceptual content (i.e.,
a linguistic description of the intended action), the set of attentional
priorities mentioned above, and a set of stimulus-response (S-R) links which are the
procedural knowledge used to execute the intended task. When an intention is
adopted it is loaded into task memory. While task memory can accommodate
more than one intention to facilitate multitasking and even modest concurrency
in execution, only one intention at a time is dominant. An intention is
dominant in task memory just in case its attentional priorities are currently guiding
the system's focus of attention. Keeping an intention dominant is a matter of
periodically focusing on it. We have implemented a facility for inner speech in
the system, such that when the name of an intention, or semantically-related
content is subvocalized, the intention remains active. Switching intentions is a
matter of a non-dominant intention receiving focus, and having its attentional
priorities guide system behavior.
Insofar as attentional priorities guide the focus to external features, regions,
objects, and events, they largely determine the content of ARCADIA's
shortterm memory stores. Items in short-term memories roughly correspond to what
Ned Block famously labeled access conscious contents, which he characterized
as being poised to be used in the rational control of thought and action [
2
].
Taken together, dominant intentions in ARCADIA partially structure (access)
consciousness in virtue of their connection to attention. We have put all of this
functionality to work in developing a simple agent corresponding to the angry
nephew in the scenarios detailed earlier. Initially, the agent has three intentions:
to plan out the speci cs of the murder, to drive to his uncles home, and
nally to kill his uncle. Because ARCADIA supports multiple active intentions,
the agent is able to drive (albeit poorly), and plan at the same time. The S-R
links associated with driving respond directly to aspects of the environment that
Rich Arti cial Agency Requires Consciousness
are processed outside of the focus of attention up to the level of coarse-grained
semantics. In the rst variant, planning dominates and the agent neglects to
periodically switch back to focusing on driving, causing the agent to unintentionally
run over what turns out to be its uncle. In the second variant, the agent is
similarly focused on planning, but a shiny red stop-sign grabs the agent's
attention, facilitating a switch back to driving. Once the agent is focused on driving,
the uncle is spotted. Recognizing the uncle a ords the agent the opportunity
to ll out the plan by using the car to run the uncle over rather than either
shooting or stabbing him, which were the two options under prior consideration.
Once planned, the agent focuses once more on driving and runs the uncle over
straightaway, killing him intentionally.
3
Final Thoughts
Along with others [
6, 5
], we recognize that any attempts to computationally
model what we ordinarily think of as consciousness will need to grapple with
perception and attention. Of course, one might imagine building a purely
Cartesian AI, but such a creature would have no appreciable connection to the external
world. Ultimately, the type of agent we have in mind exists in the world,
perceiving continuously and acting often. Selective attention structures both access
consciousness and phenomenal consciousness: when we move our eyes, peripheral
objects come into detailed focus, and the details of objects previously in focus
fade to some degree, modulo whatever detail may have been encoded in
shortterm memory. Understanding and modeling conscious cognition is ultimately a
systems-level enterprise, involving sensation, perception, attention, memory,
inference, and action. While an enormous amount of empirical work remains in
clarifying the relationships between these various capacities with respect to
consciousness, it seems undeniably true that a rich model of attention is at their
heart, and required for the kind of conscious agency that is of interest to all of
us who look to a future where machines and man are bound by the same social
and cultural mores [
1
].