Conceptual analysis and human nature 193
invariability, these are behaviours that can be imitated or simulated by a machine.
Hence, machines might be able to imitate human behaviours of this kind, but they
will not be able to – deliberately – take different routes to the same goal, as this
would require an understanding of the purposes of seemingly variant behaviors.
In the light of this argument, a machine could succeed in Turing’s imitation
game in a restricted sense, by providing an imitation of human conversational
behaviour, but it could not play that game, because playing a game would require
an understanding of the practice of playing games in general, and playing that
game in particular. However, the authors object, the machine would be unlikely to
succeed in imitating conversational behaviour in the first place, as such imitation
cannot be accomplished without an understanding of conversation in general and
the present conversational context in particular. It would be like playing Chinese
Whispers without knowing the language of the input word or phrase. Polimorphic
actions, Collins and Kusch conclude, cannot be simulated by machines, because
such a simulation would have to presuppose a genuine understanding of those
actions on the part of the machines.
Remarkably, however, Collins and Kusch use a purely ability-based definition
of human beings and machines. To them, a human being is whoever or whatever is
or could be, under appropriate conditions, capable of performing both mimeomorphic and polimorphic actions, whereas a machine is whatever is only capable of
simulating mimeomorphic actions. By adopting this definition, any hypothetical
machine that might transgress this boundary would cease to be a machine, and
the extension of the set of entities to be subsumed under the concept of human
beings would have been enlarged. Hence, the answer to Turing’s question, “Can
a machine think?” would remain negative – presuming that there is a relation
between the ability of performing actions and the presence of mental processes –
but this would be so by virtue of the machines involved losing the defining characteristics of machines and assuming defining characteristics of human beings.
One could ask whether it is an analytic statement about machines that they are
capable only of simulating mimeomorphic actions, and of humans that they are
always also capable of performing polimorphic actions, but this is what Collins
and Kusch (1998) seem to suggest, although their aim certainly is not conceptual
analysis. But then, simulating mimeomorphic actions was certainly not part of
the kinematics-based classical concept of machines, briefly summarised at the
beginning of Chapter 1, that confined the domain of what machines can do to the
movement of matter and the harnessing of energy. Even to Turing’s revision of
that view and his conception of what computing machinery can do, the imitation
of human actions was relevant only in a circumscribed sense, namely as one of the
indefinitely many possible accomplishments of his Universal Machine.
Collins and Kusch offer a conceptual escape route from one central – fallacious –
implication of the basic point of the AI sceptic that equivalence between human
cognitive and machine abilities is impossible in principle, a point the authors otherwise seem to share and which I may rephrase as follows: “Machines cannot
think because thinking is whatever a machine cannot do.” This is one possible rendering of the “moving the goalposts” machination against the notion of thinking
invariability, these are behaviours that can be imitated or simulated by a machine.
Hence, machines might be able to imitate human behaviours of this kind, but they
will not be able to – deliberately – take different routes to the same goal, as this
would require an understanding of the purposes of seemingly variant behaviors.
In the light of this argument, a machine could succeed in Turing’s imitation
game in a restricted sense, by providing an imitation of human conversational
behaviour, but it could not play that game, because playing a game would require
an understanding of the practice of playing games in general, and playing that
game in particular. However, the authors object, the machine would be unlikely to
succeed in imitating conversational behaviour in the first place, as such imitation
cannot be accomplished without an understanding of conversation in general and
the present conversational context in particular. It would be like playing Chinese
Whispers without knowing the language of the input word or phrase. Polimorphic
actions, Collins and Kusch conclude, cannot be simulated by machines, because
such a simulation would have to presuppose a genuine understanding of those
actions on the part of the machines.
Remarkably, however, Collins and Kusch use a purely ability-based definition
of human beings and machines. To them, a human being is whoever or whatever is
or could be, under appropriate conditions, capable of performing both mimeomorphic and polimorphic actions, whereas a machine is whatever is only capable of
simulating mimeomorphic actions. By adopting this definition, any hypothetical
machine that might transgress this boundary would cease to be a machine, and
the extension of the set of entities to be subsumed under the concept of human
beings would have been enlarged. Hence, the answer to Turing’s question, “Can
a machine think?” would remain negative – presuming that there is a relation
between the ability of performing actions and the presence of mental processes –
but this would be so by virtue of the machines involved losing the defining characteristics of machines and assuming defining characteristics of human beings.
One could ask whether it is an analytic statement about machines that they are
capable only of simulating mimeomorphic actions, and of humans that they are
always also capable of performing polimorphic actions, but this is what Collins
and Kusch (1998) seem to suggest, although their aim certainly is not conceptual
analysis. But then, simulating mimeomorphic actions was certainly not part of
the kinematics-based classical concept of machines, briefly summarised at the
beginning of Chapter 1, that confined the domain of what machines can do to the
movement of matter and the harnessing of energy. Even to Turing’s revision of
that view and his conception of what computing machinery can do, the imitation
of human actions was relevant only in a circumscribed sense, namely as one of the
indefinitely many possible accomplishments of his Universal Machine.
Collins and Kusch offer a conceptual escape route from one central – fallacious –
implication of the basic point of the AI sceptic that equivalence between human
cognitive and machine abilities is impossible in principle, a point the authors otherwise seem to share and which I may rephrase as follows: “Machines cannot
think because thinking is whatever a machine cannot do.” This is one possible rendering of the “moving the goalposts” machination against the notion of thinking