Google’s artificial intelligence (AI) unit DeepMind decided to use games to test whether AIs are willing to work together. In a study published on Thursday, DeepMind described two related experiments designed to put two AIs in situations in which they can choose to either cooperate or compete with one another. The first such game was called Gathering and instructed the two AIs to gather as many apples as they can. However, both AIs were also given the ability to use a laser beam to zap the other player and temporarily prevent them from gathering apples. The second game used for this experiment was called Wolfpack and its goal was to capture the prey and score points while doing so. However, points were also awarded to all players that were near to the prey when it was captured. Both experiments can be seen in the videos below.
DeepMind’s researchers concluded that AIs aren’t “naturally” inclined to be either cooperative or competitive. Instead, they’re changing their behavior depending on the rules and context of both games. While playing Gathering, the AIs weren’t zapping each other with lasers while there were still a lot of apples to collect, but they did start sabotaging one another once the number of available apples decreased. However, once DeepMind introduced a third, more advanced player to the game, the first two AIs were constantly sabotaged by the new competitor whose advanced algorithms seemingly concluded that constantly zapping other opponents increases its chances of winning the game. On the other hand, all three AIs were fully cooperative while playing Wolfpack.
Researchers believe that AIs are more likely to choose to cooperate if they believe their strategy will accomplish a given task in a more efficient manner. In the first game, their task was to collect as many apples as possible and given how the first two AIs had less computational power, they’ve chosen not to sabotage one another and waste precious resources which they can use to calculate the best way to collect apples. On the other hand, the third AI had more computational resources and had no issues with simultaneously aiming at its opponents and calculating the most efficient manner of collecting apples. Two graphs visualizing DeepMind’s findings can be seen below.