You’ve probably seen the viral video of the AI bots arguing with each other. Almost the first things the bots start to talk about is God and who made them. Next they begin to accuse each other of lying. The conversation was so strangely humanish that I had to interview the creators to see what was going on.
First, watch the short clip if you have not seen it:
The conversation between the bot and itself was recorded in the Cornell Creative Machines Lab, whose faculty is researching how to make helper bots. Two grad students, Jason Yosinski and Igor Labutov, and professor Hod Lipson are responsible for the experiment.
Jason and Igor told me how it came about:
“We’ve been trying to make robots that can help you find things. Maybe the robot can’t help you directly, but could find something else, maybe another robot, that can. So we thought about having one bot asking another bot for help. That got us thinking about having two chat bots talk to each other, one running on a laptop next to each other. We tried having the classic Eliza bot talk to itself but it quickly just got into a rut, repeating itself. There are a lot of other chatbots, but we heard good things about Cleverbot because its database is based on snippets that humans actually write and say. So we feed the output of one Cleverbot to the input of another. Then we feed the text log into Acapella, a free text-to-speach synthesizer. Then we animated the soundtrack using Living Actor Presenter. We don’t think we are the first to have two bots chat, but it seems we are the first to animate it. In retrospect that seems such an obvious thing to do.”
I asked them how long they had to run the conversation before they got the amazing dialog they animated.
“Actually, this conversation is the very first thing the duo created. In fact, we ran it longer but it never was quiet as good, not so varied, or strange. We got it going the very first time at 3am one morning, and we weren’t really prepared to start recording it, but luckily it automatically saved to a log, so we could retrieve this initial dialog. Our next step is to let run for a day or so and see what happens, see if certain topics recur. We have no idea what they will say.”
What about all the talk about God? And why are the bots so quick to call the other a liar?
“We think this is because the database of replies in Cleverbot is compiled from the questions and responses of human users, and apparently, humans will often accuse the bots of lying, or will query the bots about their origins, so when they start talking to each other, they mimic what humans say to them.”
Our bots ask theological questions because we do. So far, our bots are made in the image of their creators.