UNTIL THE Columbia tragedy of 2003, it was generally
assumed that one more space shuttle disaster would mean the beginning of
the end of manned space travel.
Indeed, no sooner had it happened than many
over-excited commentators started proclaiming that the future of space
exploration belonged solely to robotic craft. One columnist spoke warmly
of relying, no longer on astronauts, but on ``our ever more
sophisticated machines.''
But these are false prophecies, for there is an
absolute limit to how sophisticated a machine can be. As Sir Roger
Penrose showed in his superb 1989 book The Emperor's New Mind,
no machine exists, or is ever likely to exist, that matches humans in
common sense, judgement, imagination and initiative. As he put it,
``cognition is not algorithmic.'' In plain language, HAL and 2001:
A Space Odyssey notwithstanding, you cannot write software that
imitates minds that have evolved over hundreds of millions of years.
This realisation has come as a blow to
the``artificial intelligence'' community, computer scientists who have
devoted decades to trying to build a super-intelligent machine.
They understood fairly early on that there might be
two kinds of artificial intelligence; soft AI, machine behaviour that appeared
to be intelligent, and hard AI, behaviour that actually was.
Those with the more advanced versions of Windows and
access to the best websites encounter soft AI almost every day. At
lightning speed, machines do everything they have been told to do
provided they have been taught how to do it. But the drawback is that
without specific instructions they cannot do anything else.
Some people have argued that chess-playing programs
that can defeat world champions go far beyond this definition. Surely,
it has been said, such machines exhibit truly original thinking. But
this is not the case. They do not even know they are playing chess. They
are not aware of
anything at all! If the building caught fire they would calmly go on
playing. They are merely executing the algorithmic instruction:
``attempt to checkmate the king by any legal means.''
By contrast, machines programmed with hard AI would
be able to do just about anything. They could have temper tantrums, fall
in love, have a taste for bacon and eggs, terraform planets and compose
better music than Mozart.
But for good or ill, hard AI seems to be beyond the
realms of science. And even the best soft AI, because of its
limitations, makes a very poor substitute for astronauts. It may be
cheaper in cash to send only robots into space, but if we rely on them
solely we are far less likely to achieve anything useful.
I am happy that this is the case. For hard AI
machines would surely combine both the best and the worst of human
minds. They would inevitably become racist. As in many an SF story, they
would soon become irritated and frustrated by human supervision and
control. They would set out to kill or enslave us all. Or, as the AI
scientist Marvin Minsky once observed with apparent relish: ``They may
decide to keep us as pets.''
