Fatal AI mistakes could be prevented by having human teachers

Artificial intelligence needs our help. The best AIs are quickly mastering skills from lip-reading to video games, but only by learning through repeated failure. As robots take on riskier domains, like healthcare and driving, this is no longer an acceptable approach. Fortunately, a new study suggests that with the right human oversight, it might be possible to ditch the failures. To try to train an AI without it making a mistake, Owain Evans at the University of Oxford and his colleagues started with the simple two-dimensional table tennis video game Pong. Normally, a Pong-playing agent will let the ball fly past its paddle a few hundred times before realising that isn't a very good way of increasing its score.