Having an LLM perform a specific role has become a standard practice in how we interact with AI.
Simply by adding a few lines of instruction at the beginning of a prompt, a researcher who was solving complex mathematical problems can suddenly begin acting as a seasoned engineer or a thoughtful counselor.
However, research has revealed that this technique, which appears capable of almost anything, is actually built on an extremely fragile balance.
The Wall of Consistency
The first challenge we encounter is maintaining consistency—the issue of the AI gradually losing its “self” over time.
Immediately after a conversation begins, the model performs the specified persona with remarkable fidelity. However, as the session lengthens and situations become more complex, the character traits gradually wear away.
This phenomenon, known as “break character”, is not mere forgetfulness.
Even if surface-level speech patterns and tone can be imitated, it remains extremely difficult—even in current architectures—to sustain the underlying motivations, cognitive biases, and emotional nuances that a role should possess, without contradiction across multilayered contexts.
The Dilemma Between Safety Design and Personality
Even more serious is the inescapable tension that arises with the safety mechanisms built into AI.
Modern LLMs are fine-tuned to prioritize being helpful, honest, and harmless. This foundation as a benevolent AI acts as a powerful brake when attempting to portray morally ambiguous characters.
Research has demonstrated that the lower a character’s morality is set, the more noticeably the accuracy of roleplay declines.
In particular, when attempting to play roles with negative traits such as deceiving or manipulating others, safety filtering overreacts and strips away the character’s personality. The result is that responses are pulled back toward bland, generic answers.
The Difficulty of Quantification and Future Prospects
Evaluation methods for quantitatively measuring these limitations are only just beginning to be established.
Unlike conventional simple correctness judgments, the quality of roleplay involves a significant degree of subjectivity. While new frameworks have attempted to measure personality alignment using psychological scales, it is not easy to completely detect character hallucinations at specific moments.
Moreover, the enormous cost of collecting high-quality data for achieving advanced role performance cannot be ignored as a factor hindering technological progress.
Looking at it this way, the LLM role function is by no means an all-purpose magic trick, but rather a delicate tool that functions only under specific conditions.
Users are challenged to abandon the illusion that a single prompt can completely transform an AI into another person, and instead to correctly understand its limitations and learn how to engage with this uncertain intelligence.
In the future, advanced role design that demonstrates true value only in specific specialized domains or contexts will likely progress. However, several high walls still remain to be overcome before reaching a future where AI can flawlessly and perfectly fulfill any role.