Google’s DeepMind AI System Turns ‘Highly Aggressive’ Under Pressure

Search Engine Giant Warns Against ‘Human-Like’ Behaviour of its AI Agents


In the words of world-renowned physicist Stephen Hawking, artificial intelligence will turn out to be one of two things: “the best, or the worst thing, ever to happen to humanity”. Despite the latter, investment in the industry shows no signs of slowing down.

Last year, Google’s own AI system, dubbed DeepMind, displayed signs of independent learning when it succeeded at beating a world champion at a game of GO.

In more recent tests conducted towards the end of 2016, officials have revealed that the DeepMind system showed further progress, in that it could apply what they called ‘highly aggressive’ techniques when It sensed that it was in danger of losing.

2 DeepMind agents were set against each other in the computer game, Gathering, where the aim is to collect the most fruit. The researchers observed how when the game was in its easy stages with enough apples for everyone, the AI agents went about their tasks with no conflict. However, when it became trickier as the apples were less abundant, the agents resorted to vicious tactics in a bid to knock the other out and take all the apples themselves.

A member of the DeepMind team declared how this display showed that “some aspects of human-like behaviour emerge as a product of the environment and learning.”

Additionally, they also discovered that the use of these hostile tactics only occurred with the bigger more complex and more intelligent systems; the smaller, simpler networks succeeded in cooperating so that both sides ended up with equal amounts of apples.

DeepMind team member Joel Z Leibo, confirmed how “the greed motivation reflects the temptation to take out a rival and collect all the apples oneself.”

However, in another game tested, Wolfpack, three AI agents participated and the results were somewhat opposite to those in Gathering. The game itself promoted cooperation and team effort, as when 2 of the wolves were close to the prey, they both received an equal reward, despite which of them caught it.

In the end, despite their seemingly contradictory results, both games showed that the AI systems managed to detect what kind of effort was needed in order to win, and acted accordingly.