Is AI intelligent? An assessment of artificial intelligence, 70 years after Turing. (February 2022)

• Record Type: Journal Article
• Authors: Hoffmann, Christian Hugo
• Is Part Of: Technology in society. Volume 68(2022)
• Publication Date: 2022-02
• Subjects: Intelligence -- AI -- Turing -- Turing test -- Artificial general intelligence; Technology -- Social aspects -- Periodicals; 303.483
• DOI: 10.1016/j.techsoc.2022.101893
• Languages: English
• ISSNs: 0160-791X

Abstract

Abstract: 70 years ago Turing (1950, 1952), showcased his famous Imitation Game, which has come to be better known as the Turing Test. It proposed an evaluation procedure of intelligence in machines. The passage of time is perhaps reason enough to prompt the broad question: where do we stand today with regards to assessing intelligence in Artificial Intelligence (AI) systems? In this paper, we first contribute to more conceptual clarity by asking ourselves what AI and intelligence in AI is, and by comparing our answers to the latter to animal and human intelligence. We then aim to grasp the gist of the matter when we revisit Turing's proposal, criticize it, and finally inject basic requirements for a more robust and valid approach to evaluate AI systems in the future. In contrast to the standard Turing Test, which is neither valid nor robust, we propose that a measure or test of (machine) intelligence ought to lead to actionable as well as thriving research. Furthermore, the measure or test should be empirical, specific, relevant, expansive (for the specified scope), repeatable, solvable by exemplars, unpredictable, non-anthropomorphic, and, last but not least, non-binary. Highlights: Alan Turing designed a renowned test on "intelligence" 70 years ago, the imitation game which is better known nowadays as Turing Test. We cultivate a language when talking about AI that suggests that present-day machines are more human than devices; yet, in reality, current and past AI systems are more tools (namely, prediction machines) than persons. What really matters about intelligence is capacity and behavior, what is non-crucial is appearance and implementation. The standard Turing Test does not claim to be an operational Definition of humanlike intelligence. Yet, it is, unlike what Turing believes, not even a sufficient criterion for it. It just leads to many "false negatives" and many "false positives". A measure or test of (machine) intelligence ought to be non-binary, empirical, specific, relevant, expansive (for the specified scope), repeatable, solvable by exemplars, unpredictable, non-anthropomorphic, and lead to actionable as well as thriving research.