Stanford researchers identified typicality bias as human post-training evaluators rewarding artificial intelligence responses that feel familiar