Abstracts

Lessons from Artificial Intelligence on the uniqueness of human cognition

JAMES FRANKLIN University of New South Wales

Fifty years of experience in attempting to create Artificial Intelligence has cast light on where human cognitive abilities are special and where they are not. Artificial Intelligence is better understood than animal intelligence because it has been created by humans, so its achievements (and failures) provide a testbed for how well we understand different aspects of human cognition.

The extravagant promises of early AI were premised on a simple model of (human and animal) cognition, a model based on symbolic rules and search through spaces of possibilities. That model turned out to be grossly inadequate. Certain human abilities turned out to be easily imitated (indeed, easily surpassed), such as calculation and information retrieval. Of special interest for the uniqueness of human cognition are those aspects which have proved most difficult to imitate, since they pick out cognitive abilities that are not easily captured by any comprehensible model of cognition.

There are several such difficult aspects, some of which seem to be shared by animals (possibly suggesting optimism as to their eventual in-principle solution), and some not. The first concerns the transition from the vast flux of continuous information into the sense organs to the recognition of discrete objects (as in object recognition in computer or animal vision). The “middle level” of perception, between the sense organs and object identification, has not been imitated well in AI. Although animals do not speak discrete languages, they do appear to recognise and reidentify discrete objects, so this ability is not uniquely human. Nevertheless, it is important to human language; the philosophy of AI studies the “symbol grounding problem” of how to attach words to the discretized content of what they are about.

A second difficult area, which has come to the fore in areas such as formalizing legal reasoning, concerns the vagueness, fuzziness or (in legal terminology) the open texture of human concepts. That has proved hard to imitate in the standard symbolic style of AI, since symbols are discrete, all-or-nothing entities whereas human concepts expressed in the meaning of words typically have borderline cases and work in context-dependent ways. The mismatch between the two styles of concept has created unresolved problems, while casting light on the nature of human intelligence.

The most uniquely human cognitive ability, one apparently not found in animals and which there is very little prospect of imitating in AI, is understanding. Humans can not only know that 2 × 3 equals 3 × 2, but can understand why it must be so. Such human insight into necessities was highlighted in the Aristotelian philosophical tradition, but has been obscured by most contemporary schools of philosophy, to the detriment of work that aims to gain a complete picture of human cognition.

The paper surveys especially what the failures of AI has shown about the areas where human cognition is unique, with some attention to animal comparisons.