In an
interesting paper Sven Nyholm considers some of the implications of controlling
robots. I use the idea of control to ask a different question about robots and
AI more generally, in what follows the term robot will refer to robots and AI
in general. Using the word control hints at agency. We talk of driving cars,
operating machinery but of controlling ourselves, others, crowds and even our
pets. If the things we control have agency then what and how we control them matters
morally. Self-control is a virtue and controlling another competent adult a
vice. Who we control matters for humans and what we are controlling matters for
robots. The question this posting will consider is how can we know the nature
of what we are controlling when considering robots.
If a
robot is simply a tool then using a robot is an extension of our agency and
when controlling a robot we are controlling our own agency. In this context
controlling a robot raises questions as to how well we can understand our own
agency which I won’t consider here. Can a robot have some sort if agency? I
will assume here that for something to possess agency it must have some consciousness,
some might disagree. If we accept the above then when considering controlling
robots it is important to consider whether a robot can be conscious.
I am
somewhat pessimistic about the chances that a robot might be conscious and
certainly believe that none of today’s robots are. However most animals are material
creatures and appear to be conscious and it seems possible that we might be
able to construct other conscious material creatures such as robots. At the
preent time we don’t understand what makes us conscious any better than
Descartes did. It is possible that there are some things we can never know as
shown by Turing’s halting problem, we shall return to Turin later/. The nature
of consciousness might be one of these things. However most people seem to
believe we might sometime be capable of understanding the nature of
consciousness. If we can understand the nature of consciousness then it seems
probable that we could create conscious robots and that my pessimism about
doing so is unjustified. Of course such robots might be constructed from
materials other than metal and silicon. If at some future date we ask how we
should we control robots then we should first ask whether we are trying to
control a conscious or unconscious robot. This question is matters morally. For
instance how we should control and treat a sex robot would depend on whether it
was conscious or not. Accepting this of course doesn’t mean how someone treats
an unconscious sex robot doesn’t matter morally, see Inner
Virtue . The rest of this posting will consider how we might tell whether a robot is conscious or not.
Because
we don’t have a theory of consciousness it is difficult to de design a test to
decide if someone or something is conscious. If someone is in a coma we might
simply prod her but this is unlikely to work with robots because of their
construction. At this point someone might question whether we need such a test.
He might point out that by the time such a test is needed that we might have a
viable theory of conscious. In this scenario if we desire a conscious robot we
simply build it in accordance with the theory.
I accept my objector’s point but in turn point out that we don’t have
such a theory yet. Perhaps one way of building such a theory would be to
attempt building a conscious robot In this scenario we would need a test for
robot consciousness. Because of these
difficulties we might fall back on the Turing test. This basically involves
having a conversation through a terminal with a robot and if we cannot tell the
difference between this conversation and one with a person then we should
assume that the robot is conscious. If we are going to apply the test them we
assume that we can have a conversation with a robot. But must we accept this
assumption? First someone with locked in syndrome suggests the possibility that
a robot might be conscious but unable to communicate. However because we design
and build robots I find such a scenario unlikely and won’t pursue it here.
However there is another reason why a robot might be conscious and unable
converse. According to Wittgenstein if a lion could
speak we couldn’t understand him because of different ways of living. The same
argument might be applied to robots. Robots might use language we can’t
understand or inhabit a radically world. In response I would suggest that the
worlds robots and we live in touch and that any creatures that live in worlds
that touch can communicate to some degree, see if a lion could speak . Let us
accept that we might use the Turing test as a test for robot consciousness.
Unfortunately this isn’t a very reliable test as some unconscious bots seem
able to pass it. How might we sharpen it up?
In the Turing test a person tries to decide if
he is conversing with a machine or another person using a terminal and if he
cannot then he assumes that he is talking to a person. Unfortunately human
beings have a tendency to anthropomorphise things, for instance some might say it
would be wrong to kick a robot dog. It follows that human beings might not be
best placed to decide if a conversation is between two conscious creatures.
Recently Edward Tian created an app that detects essays
written by AI. This suggests that the Turing test might be amended as follows.
A person would still have a conversation with another but the decision as to
whether this conversation is between two conscious entities would be made by an
AI program. I have previously argued that general or genuine AI must have a
will, see emerging
AI . Secondly the Turing test might be amended so
the conversation with the unknown entity focusses on the will. To have a will
any entity must care about something and perhaps the conversation should focus
on what the unknown entity cares about. At this point it might be objected that
what a robot cares about might be implanted and any will resulting from this
implant isn’t genuine. In response I would argue that whilst a creatures is
defined by what it cares about, whether it is conscious or not is determined by
its reaction to this. Any creature in which what it cares about has been
implanted isn’t authentic but it can still have a will and be conscious. The Turing test might be amended so it probes
what the unknown entity cares about and how it reactions to this in different
situations. This amendment is interesting but I am doubt it would be useful at
the present time in assessing conscious but it could be a useful tool for
investigating the nature of consciousness. Lastly it might be suggested that a
Turing test might not only be assessed by AI but be conducted by some AI
system. Perhaps AI might find better questions than a human being. I would be
reluctant to accept this suggestion for it might mean accepting an attribution
of consciousness without any means of ascertaining the accuracy of the
attrition.
What conclusions can be drawn from the above?
First whilst how we collaborate with robots is an interesting question it
raises two equally interesting potential questions. First what are we
collaborating with and secondly how can we ascertain this. Both these questions
may be illusory, probably are illusory, but if they are we need to be able show
why. Trying to answer these questions is important for another reason it might
shed some light in the nature of consciousness. It also highlights the problems
of using the Turing test to answer these questions because if we accept that
animals are conscious we cannot use a Turing test to ascertain this.