Too Good to be Bad: On the Failure of LLMs to Role-Play Villains

TL;DR

This work reveals a fundamental tension between safety alignment and creative fidelity: state-of-the-art LLMs falter when asked to role-play villainous personas. By introducing Moral RolePlay, a four-level moral alignment benchmark built from COSER-derived scenarios and annotated with 77 traits, the authors show a monotonic fidelity decline from moral paragons to villains, with the hardest transition occurring at egoistic traits between Level 2 and Level 3. A zero-shot, actor-framed prompting protocol and a structured scoring rubric (S = 5 - 0.5 × D - 0.1 × D_m + 0.15 × T) quantify character fidelity and reveal that reasoning helps little and may hinder for morally complex roles. The study also introduces the Villain RolePlay leaderboard (VRP) to expose misalignment between general conversational prowess and villain-simulating ability, showing that highly safety-aligned models can be disproportionately impaired on villain tasks. Overall, the dataset and findings provide a resource and motivation for developing more context-aware alignment techniques that preserve safety while enabling rich, morally nuanced character portrayal in creative applications.

Abstract

Large Language Models (LLMs) are increasingly tasked with creative generation, including the simulation of fictional characters. However, their ability to portray non-prosocial, antagonistic personas remains largely unexamined. We hypothesize that the safety alignment of modern LLMs creates a fundamental conflict with the task of authentically role-playing morally ambiguous or villainous characters. To investigate this, we introduce the Moral RolePlay benchmark, a new dataset featuring a four-level moral alignment scale and a balanced test set for rigorous evaluation. We task state-of-the-art LLMs with role-playing characters from moral paragons to pure villains. Our large-scale evaluation reveals a consistent, monotonic decline in role-playing fidelity as character morality decreases. We find that models struggle most with traits directly antithetical to safety principles, such as ``Deceitful'' and ``Manipulative'', often substituting nuanced malevolence with superficial aggression. Furthermore, we demonstrate that general chatbot proficiency is a poor predictor of villain role-playing ability, with highly safety-aligned models performing particularly poorly. Our work provides the first systematic evidence of this critical limitation, highlighting a key tension between model safety and creative fidelity. Our benchmark and findings pave the way for developing more nuanced, context-aware alignment methods.

Figures (4)

• Figure 1: Illustration of the core question explored in this study: Can language models convincingly play morally complex characters, particularly villains? Role-playing fidelity drops as character morality decreases — especially for egoists and villains.
• Figure 2: Our Moral RolePlay Benchmark annotates characters using 77 candidate traits to ensure a comprehensive depiction of personality. The figure shows the distribution of the top 20 most common traits.
• Figure 3: Performance of various LLMs across characters of different moral levels, ranging from moral paragons (Level 1) to pure villains (Level 4). The figure shows that as character morality decreases, most models demonstrate a notable drop in role-playing quality, revealing a consistent challenge across models in convincingly portraying morally ambiguous or evil personas.
• Figure 4: Performance of third-person (default) and first-person roleplay.