AI Toy Control Problem - Stuart Armstrong (2017)

29th December 2018

A short animated YouTube video that illustrates the AI Control problem by using a simple AI agent which has a goal that is slightly misaligned with its creator's goal.

A key point is that this shows an AI with sufficient planning skills can discover deceitful behaviour which evades safety provisions. if this improves realisation of its goal, without there being any implication of self-awareness or of independent will in the agent.

The ideas are a summary of a 2015 paper and discussion on LessWrong.