While Alan Turing’s conversational test remains the gold standard for marking artificial intelligence, an annual software tournament called BotPrize has turned up a two bots which have convinced human judges that they were human. However, a Turing test involves a conversation where a person is fooled, but the winning bots from this year’s BotPrize only had to play first-person shooter Unreal Tournament.
The UT^2 bot, from the University of Texas, and Mirrorbot, developed by Mihai Polceanu, both were mistaken over half the time for humans. The four judges gave those two bots more “human” votes than the average human gamer, and over twice as many votes as any bot has ever before received. This was the first year that the BotPrize was awarded, and the two teams shared the $7310 prize.
Mirrorbot used an interesting technique to fool other players. When it met another player that it decided wasn’t a threat, it briefly “interacted” with the other avatar. The interaction consisted of mimicry: when the other player crouched, or jumped, or changed weapons, Mirrorbot did the same. Although this might appear to be standard robot behavior, it fooled the humans it was playing against. UT^2 took a different and more complicated approach: while it employed mimicry like Mirrorbot, it also deployed evolutionary learning, so it continually used more successful strategies and appeared more human-like.
Although the BotPrize is considered to be a lesser challenge than the Turing test, and the winners aren’t claiming that their bots are intelligent, it’s still a huge milestone for artificial intelligence. Although I am surprised that no team entered a bot that simply trolled opponents in a preteen voice. That would’ve fooled everybody used to playing online shooters.