Researchers from Tencent and Sun Yat-sen University have found that built-in safety mechanisms prevent large language models from effectively roleplaying negative characters, replacing subtle manipulation with generic aggression.
