Google's DeepMind team has just detailed one the projects it's working on that revolves around that exact subject. They have also created two demos (seen in the videos below) to showcase how the AI behaves in different scenarios.
In the above demo, called "Gathering," two AI entities (red and blue) work independently to gather apples (the green blobs), and when there are plenty of them scattered about, peace is achieved. But as soon as the supply begins to dwindle, the aggressiveness of each AI ramps up, and the two effectively duke it out.
What's interesting here is that each AI is equipped with a laser, which is used when the going gets tough, so one AI tries to keep the food from the other. It's a simple model to observe, but emulates a critical real world situation living condition, unfortunately.
In the other demo, called "Wolfpack," two red AI work together to capture prey while navigating the environment. In essence, the more these two AI work together, the greater the chance of success, not just for capturing, but for protecting the their food resource after the hunt.
While the findings of these tests are admittedly not that surprising, the fact that the conclusions were reached entirely through deep learning is what's really impressive. Through complex math, these AI had to learn on-the-fly how to best approach the problems they had to solve, with no human input. And if that means trying to fight the other AI off so that it can stock enough food, that's just what's going to happen. Conversely, learning to cooperate is obviously a highly valuable concept as well.
Survival mode scenarios are of course a necessary skill exercise for any being, human or AI. And though being paranoid of a Skynet uprising is a healthy counter-balanced perspective as AI becomes more powerful, clearly our robot overlords are still in their infancy.