AI has a dangerous bias problem -- here's how to manage it