튜링 테스트 재조명과 최신 비판 여섯 가지

📝 Abstract

Considering that Turing’s original test was co-opted by Weizenbaum and that six of the most common criticisms of the Turing test are unfair to both Turing’s argument and the historical development of AI.

💡 Analysis

Considering that Turing’s original test was co-opted by Weizenbaum and that six of the most common criticisms of the Turing test are unfair to both Turing’s argument and the historical development of AI.

📄 Content

In Defense of the Turing Test and its Legacy∗ Bernardo Gon¸calves National Laboratory for Scientific Computing (LNCC), Brazil goncalves@lncc.br November 20, 2025 Abstract Considering that Turing’s original test was co-opted by Weizenbaum and that six of the most common criticisms of the Turing test are unfair to both Turing’s argument and the historical development of AI. The Turing test has faced criticism for decades, most recently at the Royal Society event “Celebrating the 75th Anniversary of the Turing Test.” The question of the Turing test’s significance has intensified with recent advances in large language model technology, which now enable machines to pass it. In this article, I address six of the most common criticisms of the Turing test: • The Turing test encourages fooling people; • Turing overestimated human intelligence, as people can be easily fooled (the ELIZA effect); • The Turing test is not a good benchmark for AI; • Turing’s 1950 paper is not serious and/or has contradictions; • Imitation should not be a goal for AI, and it is also harmful to society; • Passing the Turing test teaches nothing about AI. All six criticisms largely derive from Joseph Weizenbaum’s influential reinterpretation of the Turing test. The first four fail to withstand a close examination of the internal logic of Turing’s 1950 paper, particularly when the paper is situated within its mid-twentieth-century context. The fifth also arises from a naive view that overlooks the political-economy context and how AI is imbricated in the broader history of automation in modern societies. The sixth also arises partly from “the AI effect” (discussed below) and partly from the dynamics of polarized debates. Does Turing’s original test encourage fooling people? Turing started using the term imitation after his wartime experience breaking the Enigma codes by studying the machine’s behavior alone. If its actions could be predicted with computing machines, he wondered, might computers not be capable of “imitating” the brain? Turing’s proposed question was whether a machine could learn to “play the imitation game so well” that it would be mistaken by ordinary people for what it is not. He framed the problem this way because some of his closest critics insisted that machines would never be capable of mastering language, a capacity they viewed as the hallmark of thinking (Gon¸calves, 2024b). A decade later, Weizenbaum shifted the focus of this debate from the possibility of machine intelligence to the (no less important) issue of human susceptibility to being misled by computing technologies. ∗This article develops the author’s contribution to the session “The Turing Test and its Legacy, 1950-2025” at The Next Turing Tests Conference, held in October 2025 at the University of Cambridge. 1 arXiv:2511.20699v1 [cs.CY] 24 Nov 2025 Weizenbaum’s question appeared as the title of a paper he published in 1962 in the Datamation magazine, “How to Make a Computer Appear Intelligent” (Weizenbaum, 1962, his emphasis). At first sight, Turing’s and Weizenbaum’s questions may seem remarkably similar. To see the difference more clearly, note that Weizenbaum had already achieved his main result by 1966 (Weizenbaum, 1966) — just four years later. Turing, by contrast, envisioned a fifty-year project that ultimately took more than seventy. These differing time frames are revealing. Turing sought to explore whether machines could learn enough to imitate what was widely accepted as human intelligence. Weizenbaum, by contrast, bypassed the learning question altogether: he set out to construct the simplest possible computer, pre-scripted with psychological prompts, to test whether people would attribute intelligence to it. Did Turing exclude Weizenbaum-style machines from his test? Yes. This is evident, for example, in his discussion of the “human fallibility” he encouraged the machine to exhibit. Fallibility was intended to emerge as a by-product of learning from experience, rather than from trickery: Another important result of preparing our machine for its part in the imitation game by a process of teaching and learning is that ‘human fallibility’ is likely to be omitted [from the teaching] in a rather natural way, i.e., [learned] without special ‘coaching.”’ (Turing, 1950, p. 459) A machine qualifies for Turing’s test only if it was not engineered specifically to pass it — a condition that excludes Weizenbaum’s interpretation of the test. Otherwise — and this is a crucial point — the situation would be akin to a physicist having to check for sabotage before running an experiment. Further clarification came in Turing’s 1951 lecture, “Can Digital Computers Think?”: This [learning] process could probably be hastened by a suitable selection of the experiences to which it was subjected. This might be called ‘education.’ But here we have to be careful. It would be quite easy to arrange the experiences in such a way that they automatically caused the structure

This content is AI-processed based on ArXiv data.