A Short Exploration of How AI Might Become Evil

This video is a short exploration of where we might go wrong with AI. The idea of a utility function represents what an AI "wants" to do, according to how it's designed, and misunderstanding the implications of maximizing a certain utility function in an unexpected way is a real danger. The video also touches on the idea of boundaries, where our expectations of the context of an AI's actions might not align with reality, which can have disastrous and far-reaching implications. This is a pretty well-explored subject in science fiction - HAL 9000 and other common examples spring to mind. Another example might be a robot security guard protecting a company that might decide to not let the owner in if it perceives he is damaging the company with his business decisions, but he might also impersonate him on social media with a deepfake video and ruin his reputation in order to get him fired. Our expectations for how utility functions might be interpreted can be completely subverted, and intent plays no role in how AI might act.

Search the Internet for "Control Problem" to learn more about this important subject.