Google DeepMind researchers have found a new way to make use of data deemed unsafe for AI training. Labs try to avoid data that is toxic, inaccurate, or contains personally identifiable information.