if you can view others as being like yourself, then
you can imitate them by mapping their actions onto
your own possible movements and sounds. Autistic
children are taught to imitate non-verbal behaviour
before the teachers will attempt to build speech (Lo-
vaas, 1977). Pepperberg’s (1999) grey parrot was
able to learn much more quickly and flexibly than
previous classical conditioning attempts through a
technique called model-rival learning. Rather than
learning the objects and associated words directly,
the learner observes someone else learning the task
and can then imitate them.
In Savage-Rumbaugh’s (1994) experiments involv-
ing pygmy chimpanzees who typed symbols on a key-
board, it was the chimp’s child that learned best
from observing her mother being taught. Kanzi was
eventually able to comprehend complex and novel
sentences such as, ”Kanzi, if you give Austin your
monster mask, I’ll let you have some of Austin’s ce-
real”. The evidence that he understood the mean-
ing is that he gave Austin the mask and pointed at
his cereal. Kanzi was also able to demonstrate his
comprehension without expecting immediate reward.
Pepperberg and Savage-Rumbaugh both found that
consistent reinforcement of the meaning of a word
was more effective than a simple food reward on cor-
rect production. If the pupil says ’hat’ on seeing a
hat, it was better to let them touch the hat, then
give them a treat.
An early stage in teaching children with deafblind-
ness is to demonstrate to them that their actions can
specifically affect others. A machine must also realise
this, and make the appropriate predictions about the
outcomes of it’s behaviour. Only then can we make
the correct associations between what was said and
what happened.
Seeing others as intentional beings also makes
them more interesting. Other creatures can be a
source of information, unpredictability and mutual
understanding. Pepperberg and Savage-Rumbaugh’s
teaching methods rely heavily on the subject en-
joying social interactions. The parrots did not
learn nearly as effectively when they were shown
videos of the objects without human involvement
(Pepperberg, 1999). Therefore a degree of sociability
may be important.
6. Sharing Sociability
It seems there is a bias in us for paying more atten-
tion to human sounds, shapes and movements, and
when this malfunctions, as in autism, communica-
tion can be limited and often non-existent. Severely
autistic children appear to view other humans sim-
ply as ob jects. This clearly makes engaging the
child in interaction extremely difficult, as the strange
noises coming from the bags of skin are not regarded
with any more interest than other ambient sights or
sounds. This highlights the need for the machine not
only to be able to share attention with the human,
but also to find interacting interesting. This can be
as simple as it concentrating on human skin colour
and noises from the frequency spectrum of human
speech (Breazeal, 2002), but this is by no means an
open-ended solution.
Sociability also involves the gestures, facial expres-
sions and other sounds that make up a large part of
human and animal communication. To encompass
this additional information there has been recent in-
terest in equipping embodied agents with similar so-
cial skills (Breazeal, 2002). Sociable machines hope
to make the interaction more natural for humans and
allow us to use our highly developed abilities for in-
terpreting and manipulating others. Humans uncon-
sciously project intention onto autonomous agents
(Reeves and Nass, 1996), and will even manipulate
their utterances to achieve a more satisfying interac-
tion. It may be possible to exploit these tendencies
for the benefit of both.
7. Sharing Consciousness
(Wenburg and Wilmont, 1973) define communica-
tion as any attempt to achieve understanding. But
the possibility of an understanding machine is still
widely disputed. (Harnad, 1990) has made clear that
the only understanding that currently takes place in
our computer programs, is when a human mind in-
terprets the information. A machine manipulates ab-
stract symbols purely by their shape, and attributes
no meaning whatsoever to them. The feeling is that
input and output behaviour is not enough to prove
the machine understands (Searle, 1980). In the case
of a machine, translating my email into Chinese, does
not mean it understood it in English.
Human-animal communication research suffers
from similar criticism. The learning of correct re-
actions for sounds, pictures or gestures can be ex-
plained by classical conditioning and does not nec-
essarily demonstrate the possession of meaning or
understanding. Early attempts used food rewards
as the primary motivation for the animal. But
(Savage-Rumbaugh and Boysen, 1980) admits that
chimps can easily acquire skills that appear like lan-
guage on the surface, without demonstrating the use
of symbols or other important interaction behaviour.
”Knowing how to use the symbol ’banana’ as a way
of getting someone to give you a banana is not equiv-
alent to knowing that ’banana’ represents a banana.
p. 67” (Savage-Rumbaugh, 1994) Similarly, children
with deafblindness can be taught segmented and dis-
crete skills that may help them get what they want
in certain situations without shared meaning. But
there communication development is stuck at a basic
stage and they will not learn to communicate fluently
and flexibly.
113