Book Reviews

Smith, Brian Cantwell. The promise of artificial intelligence: reckoning and judgement. Cambridge, MA: MIT Press, 2019. xx, 157 p. ISBN 978-0-262-04304-5. £20.00/$24.95.

I doubt if anyone will be surprised by the author's central point, expressed in the first sentence of the introduction:

Neither deep learning, nor other forms of second-wave AI, nor any proposals yet advanced for third-wave, will lead to genuine intelligence. Systems currently being imagined will achieve formidable reckoning prowess, but human-level intelligence and judgement, honed over millenia, is of a different order.

This is not to argue that artificial intelligence has not, or will not, lead to considerable, beneficial developments: already, Google translate is much better than it was (for some languages) as a result of its new machine learning methods; and the employment of machine learning statistical techniques operating on enormous data stores is likely to lead to new discoveries of the relationships, for example, between diseases and genetics. General, human intelligence, however, involves more than rational thought and computation: as the author notes, it involves judgement and judgement involves cultural norms, ethics, and practices, and what is appropriate in different situations. We make judgements about what right, fair, and proper - concepts that are difficult to embed in systems, whatever their reckoning power.

We see the hazards of reliance upon the reckoning power of artificial intelligence systems in the six fatalities so far reported for self-driving cars, five of which involved the Tesla electric car. True, some of these appear to have been the result of drivers who were competing for a Darwin Award, but others could have been avoided if human judgement, rather than simple computation, been built into the program.

The author pursues his initial statement through chapters dealing with the history of artificial intelligence from the first phase, referred to as GOFAI (Good Old Fashioned AI) to the present, with a look at proposals for a third-wave variety. Part of the problem, implicit in the self-diving car problem, is that humans are capable of interacting with the world not only through their senses but also in the light of previous knowledge, cultural mores, context, and more. Systems are only capable of interacting whatever terms have been built into the program by the system designer. So far, we do not have self-aware systems that could act independently in the world.

The author concludes that, regardless of what has been learnt from the failure of GOFAI, and from the second-wave of articicial intelligence, based on machine learning, he does not see:

anything on the horizon—in scientific or technological or even intellectual imagination— that suggests we are about to construct, or are even thinking about constructing, systems capable of full-scale judgement

which would result, for example, in 'systems that will go to bat for the truth, reject what is false, balk at what is impossible—and know the difference'. We would also have to ask whether any of today's tech giants would want to develop such a system!

This is a well written and readable book on an abstruse and difficult (philosophically) topic and the author rightly takes computer science to task for giving well understood words different meanings within the discipline. He mentions specifically, semantics, reference and interpretation, and certainly the misuse of semantics as in referring to certain HTML tags as 'semantic' does grate! However, the author does occasionally adopt terms from other disciplines and use them in a more general sense: I had to look up, for example, distal, which is used as a synonym for 'remote'; and imbricated for 'overlapping', and it really cannot be said that ideas overlap in the same was as do the scales of a fish, or roof tiles.

Professor T.D. Wilson
Editor-in-Chief
December, 2019

How to cite this review

Information Research is published four times a year by the University of Borås, Allégatan 1, 501 90 Borås, Sweden.