Requiring a human being to be on the public record as personally and financially responsible for an AI's decisions would be a good start.