AI System Behavior

Teaching AI Systems to Behave Themselves
The New York Times  by Cade Metz

A small but growing community of artificial intelligence (AI) researchers is beginning to explore mathematical techniques that aim to prevent machines from demonstrating unexpected, unwanted, and harmful behavior. Safety concerns are growing as AI moves into online services, security devices, and robotics. Researchers at OpenAI are developing algorithms that can learn tasks through hours of trial and error while receiving regular guidance from human teachers. The researchers believe algorithms that combine human and machine instruction can help keep automated systems safe. OpenAI is working to build reinforcement-learning algorithms that accept human guidance to ensure that systems do not stray from assigned tasks. Some researchers in the field are working to ensure that systems do not make mistakes on their own, while others aim to prevent hackers and other bad actors from exploiting AI flaws.  ,,, "