I recently finished a fascinating book by Brian Christian, The Most Human Human. Christian took part in the 2009 “Loebner Prize,” an annual “official” Turing Test, in which prizes are given to the program which is able to fool the most human judges (or elicit, on average, the lowest confidence of its non-personhood in the judges’ assessments). Christian did in fact win the “Most Human Human award,” and to sum up some of his insights on how far artificial intelligence has come in terms of holding a conversation, I crafted a short Citizen’s Guide to Judging whether it is a Computer or Not:
1) Don’t have an argument with it: Christian mentioned that programs designed to be aggressive, hostile, and confrontational are most likely to fool a human instant messaging with it. He mentions an anecdote of a college student who had inadvertently gotten caught up in an online argument with a computer and had gone on for over an hour (although he did note 20 minutes in that the opponent “sounded like a goddamn robot.”) Have you ever had an argument where you attempt to follow a careful line of reasoning, but with every sentence, your opponent unreasonably clings to some phrase or something you said and takes it WAY out of context? It must happen to everyone. All of a sudden you have a conversation that is not reliant on anything that was said a long time ago, nor even on the actual meaning of what you said. All the computer has to do is to identify a couple keywords from your response, run a matching search on a database of some inflammatory phrases, and fire one back at you. A lot of chatbots, Christian mentions, don’t even retain memory of more than 3 previous lines of the conversation – unnecessary in an argument in which neither side is really listening to what the other is saying.
And thus, explains Christian, heated arguments are among the “least human” things we humans do – least indicative of thought development, originality, self-reflection, and empathy. In fact, an accurate measure of how “human” an activity is could easily be how difficult it is to write a program that is indistinguishable from a human. And on that scale, heated arguments rank pretty low.
2) Ask it the same question repeatedly: As mentioned earlier, chatbots don’t normally hold memory of the entire exchange, only a few of the previous lines, if even that, so they wouldn’t “notice” anything weird going on. A second reason is, Cleverbots, conversationalist programs that have been successful in recent Loebner Prize competitions, build up their databases from human responses in former conversations so that their responses seem as human as possible. They easily fool people but they easily contradict themselves. One exchange showed a cleverbot contradicting its own gender 3 lines apart in a conversation. One thing to understand is, against a program, you’re not really against “one person” or thing, but rather many. Similarly, Deep Blue’s victory over Kasparov wasn’t really against “one chess player” since decades of AI-chess expertise and advice and tutelage from several grandmasters had gone into the program.
So if you ask a human “How are you?” 10 times in a row, he or she would probably start to get annoyed or confused, and express that. If you ask a chatbot “How are you?” 10 times in a row, “he or she” would probably give 10 identical answers, unperturbed, or just as bad, 10 different answers.
3) Don’t let it dominate the conversation: Christian offered an example of a “Most Human Computer” winner from the late 90’s whose only topic of conversation was a White House event involving Ellen DeGeneres and Bill Clinton (no sex scandals that time). Judges who engaged the bot were easily fooled, and those who knew nothing about it or changed the topic halfway through saw it get confused and nonsensical. This can backfire; Christian noted one British man who was asked a question about Sarah Palin and responded “I’m sorry, I don’t know who that is.” I can’t remember, but I don’t think he won Most Human Human.
So there you have it – an introduction to where AI has gotten in terms of the Turing Test.
Here’s a Daily Show interview with the author, Brian Christian:
Christian emphasizes literature and art about how to be as human as possible. We see machine-like behavior in humans everywhere – think of generic text message conversations, or even chess grandmasters (nary a grandmaster does not memorize opening and closing tactics). For my own part, in studying mathematics, I subconsciously move away from mechanistic practice (if you’ve solved the three types of system of linear equations with 3 unknowns, you’ve solved them all) and more towards what a computer couldn’t do – translating real-life problems into the realm of mathematics, where a computer may or may not be able to do the job.
But literature and the arts are in fact the most difficult to program. So engineers and scientists, don’t be cocky – you’ll be the first to lose your jobs when the robots inevitable take over.
“Computing Machinery And Intelligence” – Alan Turing (I’ve cited this before, but it’s ever-accessible and fun to read and re-read)