Unreal Tournament programming competition creates AI bots that are more convincingly human than actual humans

What a sad day for the Human Race. Turns out we can't even win at 'being human' any more. Still, chin up. Let's look at how this shameful state of affairs came about.

It's all thanks to Unreal Tournament 2004 or, more specifically, a competition at the IEEE Conference on Computational Intelligence and Games, which used the framework of Unreal Tournament 2004 to create Artificial Intelligece in bots that could trick human players into thinking they were playing against fellow sacks of carbon and water (according to Eurogamer).

Normally, humans are human enough to be distinguishable from AI bots but, for the first time ever, it was the AI bots that came out on top in a 'Turing test' last week. In case you haven't heard of a Turing Test before, it's a 100-year-old method designed to test a machine's apparent consciousness, by judging its ability to fool us into thinking it's human. The aim was to beat the 50% threshold and that was achieved for the first time in this expeiment, with a score of 52%. Most humans only score around 42%, so that's pretty remarkable.

Above: You'd think the judges could just ask something over voice chat, wouldn't you?

Funnily enough, this breakthrough was actually achieved by making the bots a little more rubbish at things, such as aiming and tactical decisions. They were given new parameters for illogical human traits like revenge - chasing after one player even if it meant jeapordising their own safety. Aiming accuracy after fast movements was also toned down, as was the ability to hit targets at long distance. Basically, they were noobified.

With two bots being tagged as 'human' the same number of times by the judges, the $7,000 prize was split between two parties: UT^2, created by University of Texas professor Risto Miikkulainen alongside doctoral students Jacob Schrum and Igor Karpov, and MirrorBot by Romanian computer scientist Mihai Polceanu.


  • Imgema - September 28, 2012 1 p.m.

    This article reminded of Perfect Dark. There was the "Revenge Sim" (they were called sims instead of bots) that would try to kill the player or sim who killed it last time...
  • ColaFlavourChewits - September 28, 2012 9:23 a.m.

    Interesting read. It seems AI is too often a neglected part of the development cycle of modern games; even massive titles like Skyrim rarely get beyond 'draw weapon and run at player' if it's a melee character or 'stay back and shoot' if they have a bow. Particularly irritating is when you drain a bandit's health and they fall to the floor screaming "I yield!", then after a few seconds get up and start attacking again! (Or maybe the bandits of Skyrim are optimists by nature) The article also highlights something that may in fact be more important than intelligent AI for immersion; fallibility. It's great that enemies flank, find cover and pick their battles, but they seem to be the most world-weary enemies the world has ever seen; in Crysis a man in a robotic suit leaps from the bushes onto a building while carrying a soldier by the throat in one hand, and everyone just turns and says "Get him!". Wouldn't it be nice if sometimes the soldiers said "Sod this, no way I'm fighting a robot man." and legged it? If I can't have that, then at the very least I want them to miss once in a while when shooting at me from a mile away through a forest. Unless Korean soldiers are that good, in which case I rescind my complaint.

Showing 1-2 of 2 comments

Join the Discussion
Add a comment (HTML tags are not allowed.)
Characters remaining: 5000