If you though that George from Jabberwacky was cool, wait until you see the chatbots in Second Life. They display facial expressions and other forms of nonverbal communication, they can navigate and interact with their virtual environment through script and sensors, all while carrying on the same natural language conversation used by web based chatbots.
George that I mentioned earlier has gained recognition for winning the 2005 Loebner prize for being the most advanced artificial intelligence on the planet. The Loebner prize is based on the Turing Test, don’t worry we’ll get to that in a moment, and there seems to be a heated contest between Richard Wallace’s A.L.I.C.E. and Rollo Carpenter’s George for the title. Although Carpenter’s winning entry this year was named Joan. Still, these men have devoted amazing research and time into developing chatbots to compete in this test. What is interesting to me, and what I think these gentlemen have figured out, is that the Turing Test is not about producing human level intelligence. It is about fooling human level intelligence into not noticing that they are discoursing with a program that is not as smart as them.
What do I mean by that you might ask. Combing through AIML, Artificial Intelligence Markup Language which is an amazing program concept in and of itself, it becomes apparent that there are only so many decision trees that you can account for in a limited space. Even when, as discussed in the literature at the A.L.I.C.E. foundation’s website, many conversations lead to the same branches, it is the random quirkiness of human interaction that determine our ability to pass in what is primarily a highly evolved society of sensory apparatus. Much of our energy consumption is used to support that grey goop sitting on your neck, and much of that is devoted to silly things like communication, social interaction(yes it is social even if it is just online), and the root of all intelligence – abstraction. To get back to the point, the best chatbots use subtle types of distraction when the conversation goes beyond the boundaries of there main decision trees to direct things back to a manageable point. They throw off-hand comments, or ask seemingly random questions to pull you back into the illusion. It is a great tactic, because it is one that human’s employ frequently, and it can be mistaken for the quirkiness of intelligence. Also it is good for a few laughs, because let me tell you, those bots can kick out some zingers when your not expecting it. But what does this all mean for our search for intelligence. I’m not sure that is the point of the exercise.
Alan Turing set out very explicit terms in his 1950 paper "Computing machinery and intelligence" to find something that could imitate a humans conversation so well that you couldn’t tell the difference. Fake, deceive, pull one over. The test is to mislead a human judge into thinking the computer is the human in a three person conversation. At the time he suggested that a teletype be used to limit the test to linguistic capabilities, because the technology at the time was insufficient to bring other factors such as an audible voice into the equation.
So this brings us finally to the thing I really want to talk about, chatbots in Second Life. If you haven’t heard about Second Life, I’m sorry. Put the WOW down for a minute, check out what’s going on, because the potential with SL is the closest thing we have seen to the Metaverse, by a long shot. Second Life is a 3D world created and maintained by the residents. Stop the presses! The users, the masses, the inmates, are running the asylum. Objects are created, and I talking about amazing 3D renderings of whole worlds, and endowed with awesome functionality through a very straight foreword scripting language.
Chatbots, using the llHttpRequest function, are beginning to show up in Second Life. People are interested in them to add interaction to there casino or dance hall. Serve as salesbots in their stores, bartenders in there clubs, and educational institutes are looking towards them as information providers and basic instructors. What makes me intrigued is there ability to expand the concept of the Turning test. Turning chose to limit the concept at the time to text based communication, because it would be too hard otherwise. I see his point, but I think we are missing something. Yes, trying to mimic voice patterns that express emotion is incredibly hard. Adding body movement in a 3D environment just seems like making the test harder. But so much is added with the ability to include nonverbal communication. That subtle some what random shift of the human body with whom we are conversing reassures us is ways we don’t consciously recognize. Things like posture and other hints like facial expressions provide enrichment that is lacking in text limited conversations.
Not to mention that amazing potential of chatbots that can easily draw upon environmental information, as well as control that same environment as well. Using scripts the chatbots can access location and status of objects and avatars around them, and access any and all functions we can think to add. For example, I can conceive of bots that can control video feeds, manipulate object size, location, or function. Bounce unruly guests, and fetch needed tools.
The addition of sensory information, and environmental interaction, greatly enhance the capabilities of AI in ways that we hope to see with robotics here in our first life over the next few decades. It adds depth that web based chatbots will never accomplish. It could go a long way in fooling you into thinking it might be controlled by human level intelligence, and isn’t that what we all are trying to do. Okay, maybe not the human part, but certainly the intelligence.
