Sonje Finnestad

Given the difficulties of defining “machine” and “think”, Turing proposed to replace the question “Can machines think?” with a proxy: how well can an agent engage in sustained conversation with a human? Though Turing neither described himself as a philosopher nor published much on philosophical matters, his Imitation Game has stood the test of time. Most understood at that time that success would not come easy, but few would have guessed just how difficult engaging in ordinary conversation would turn out to be. Despite the proliferation of language processing tools, we have seen little progress towards doing well at the Imitation Game. Had Turing instead suggested ability at games or even translation as a proxy for intelligence, his paper might have been forgotten. We argue that these and related problems are amenable to mechanical, though sophisticated, formal techniques. Turing appears to have taken care to select sustained, productive conversation and that alone as his proxy. Even simple conversation challenges a machine to engage in the rich practice of human discourse in all its generality and variety.

Showing remarkable insight into the relationship between language and thought, Alan Turing in 1950 proposed the Imitation Game as a proxy for the question “Can machines think?” and its meaning and practicality have been debated hotly ever since. The Imitation Game has come under criticism within the Computer Science and Artificial Intelligence communities with leading scientists proposing alternatives, revisions, or even that the Game be abandoned entirely. Yet Turing’s imagined conversational fragments between human and machine are rich with complex instances of inference of implied information, reasoning from generalizations, and meta-reasoning, challenges AI practitioners have wrestled with since at least 1980 and continue to study. We argue that the very fact the Imitation Game is so difficult may be the very reason it shouldn’t be changed or abandoned. The semi-decidability of the game at this point hints at the possibility of a hard limit to the powers of technology.

In 2014, widespread reports in the popular media that a chatbot named Eugene Goostman had passed the Turing test became further grist for those who argue that the diversionary tactics of chatbots like Goostman and others, such as those who participate in the Loebner competition, are enabled by the open-ended dialog of the Turing test. Some claim a new kind of test of machine intelligence is needed, and one community has advanced the Winograd schema competition to address this gap. We argue to the contrary that implicit in the Turing test is the cooperative challenge of using language to build a practical working understanding, necessitating a human interrogator to monitor and direct the conversation. We give examples which show that, because ambiguity in language is ubiquitous, open-ended conversation is not a flaw but rather the core challenge of the Turing test. We outline a statistical notion of practical working understanding that permits a reasonable amount of ambiguity, but nevertheless requires that ambiguity be resolved sufficiently for the agents to make progress.