As an adolescent, I invented a silly landmark for myself. I would only consider myself proficient in a language if I could “get” a joke in that language which had no straightforward translation into my primary language.
Hardly a scientific milestone, but it seemed to make sense back then, since “getting” a joke requires at least sufficient vocabulary and grammatical training to understand whatever the joke-teller was saying.
I’m wondering whether this silly signpost could be flipped around to evaluate Artificial Intelligences. After all, AI is at the convergence of natural language and computational thinking, and the Turing Test is a well-known measure of an AI’s success at human-like intelligent communication. In brief, the Turing Test is a theoretical experiment where a “blind” human observer has a conversation with an AI and another human. If the observer can’t tell which of two is actually human, the AI has successfully passed the Turing Test.
My proposed extension to the Turing Test is simply to make certain that the conversational exchange involves jokes being told by all participants. The quality and nature of the humor in the banter might be strong indicators of a given speaker’s humanity. Jokes require basic linguistic skills, of course, but they also require more advanced contextual knowledge. Many forms of humor rely on the reversal of expectations, whether in the context of the narrative or in the meta-context of the language being used to convey the narrative. What’s more, I speculate that most truly funny jokes involve both linguistic “wordplay” as well as in-narrative reversals.
Since successful human communication requires understanding both text and context, and “getting” jokes clearly necessitatex such an understanding, the Turing-Ideambulate Modified Test might be a more rigorous and accurate test of human-like artificial intelligence.
Plus, it would certainly make life more fun for the judges at the Loebner Prize.
