The Turing test has long been a major tenet of machine-human interaction and the study of Artificial Intelligence. Devised in 1950 by Alan Turing, a British informatician, computer scientist, mathematician and code-breaker who famously helped crack the German ENIGMA code during WWII, the test attempts to address a long-standing and thorny question in AI theory: “Can machines think?” Turing’s proposed test was genius in its simplicity: Judge the bot against the gold standard on this planet — human beings. Turing suggested that if a bot could successfully convince 30 percent of human judges that it was itself human, it would be truly “intelligent” artificial life.
No bot to date has passed the test, but they’re getting closer. This September, during Turing’s centenary year, a team of computer scientists here at UT led by Professor Risto Miikkulainen claimed the BotPrize. The BotPrize, brainchild of Australian AI researcher Philip Kingston, is a modified version of the Turing test that uses video games as a medium. The goal of the competition was to successfully convince human players in a first-person shooter video game that their opponents were also human when they were in fact robots. The UT program performed incredibly well, “out-humaning” the humans by achieving a convincing 52 percent “humanness rating” compared with only 40 percent for actual humans.
Admittedly, even the most sophisticated bot in a video game is limited to activities far more simplistic and constrained than most human socializations, but on some level, playing a video game is a subset of the same human interaction that the Turing test mandates is the ultimate benchmark for intelligence. These bots didn’t just approach this benchmark, they surpassed it.
This has me excited. Not because I want a new breed of robot overlord or want to become god in the machine myself, but because research like this allows us to probe not only the fascinating field of machine-human interaction, but also the way we interact with other humans. The strange thing about video game bots is that it’s easy to create a bot that is a force to be reckoned with, that dominates human opponents on the virtual battlefield. But it’s very hard to mimic real humans convincingly, because humans aren’t perfect. They’re inconsistent. They hold grudges and make patently unwise emotional decisions. They rage and cheat and calculate and scheme and fail. And if we can learn to code all these flaws and biases and quirks, then maybe we can examine them, better understand them and even fix them. The Turing test and the BotPrize aren’t just important ways of measuring intelligence. They’re also compendiums of those things which we humans consider important, and those flaws which are most evident. They’re not just metrics by which we judge a robot’s humanity, they’re metrics by which we judge our own as well.
McBee is a biology junior from Austin.
