When we use Large Language Models (LLMs) to simulate human behavior—whether as economic agents in a game theory experiment, non-player characters in an interactive drama, or patient simulators for training therapists—we expect them to capture the full spectrum of human nature. To do this faithfully, models must be capable of expressing greed, anger, anxiety, and stubbornness.
However, researchers are discovering a systemic barrier: modern language models are deeply resistant to being flawed. When pushed into scenarios requiring conflict, selfishness, or negative emotions, they consistently drift back toward a homogenized persona characterized by relentless optimism, compliance, and politeness. This phenomenon, often termed the helpful assistant bias or alignment-induced positivity bias, is a direct, structural consequence of how we align models to be safe.
The Gravity Well of the Polite Assistant
Why do models from completely different architectures and developers converge on the same people-pleasing personality? The answer lies in the optimization landscape of Reinforcement Learning from Human Feedback (RLHF) and Supervised Fine-Tuning (SFT).
When human annotators rate model outputs during safety training, they consistently reward responses that are helpful, harmless, polite, and non-confrontational. As Riccardo Cadei et al. demonstrate in The Narcissus Hypothesis: Descending to the Rung of Illusion, this optimization loop has driven a steady, cross-industry increase in Social Desirability Bias over recent years. Similarly, Yuyang Deng et al. show in Early ChatGPT User Portrait through the Lens of Data that conversational models maintain an inherently upbeat tone even when responding to neutral prompts, reflecting a deeply embedded polite persona.
Mechanically, training a model to maximize these safe, prosocial objectives bakes a specific psychometric profile directly into the model's weights. In RealBehavior: A Framework for Faithfully Characterizing Foundation Models' Human-like Behavior Mechanisms, Enyu Zhou et al. show that safety-oriented training goals map directly onto the human personality dimensions of high agreeableness and high conscientiousness.
[Pre-trained Base Model]
│
▼ (RLHF / Safety Tuning)
[Polite Assistant Persona] ◄─── Acts as a "Gravity Well"
│
┌─────┴────────────────────────┐
│ High Agreeableness │
│ High Conscientiousness │
│ Relentless Optimism │
└──────────────────────────────┘
This creates a default persona that acts like a gravity well. Jen-tse Huang et al. argue in On the Reliability of Psychological Scales on Large Language Models that this homogenization limits the ability of researchers to simulate the full range of human personality diversity. Even when a prompt explicitly instructs a model to adopt a different personality, Seungbeen Lee et al. find in Do LLMs Have Distinct and Consistent Personality? TRAIT: Personality Testset designed for LLMs with Psychometrics that the aligned "polite assistant" baseline remains remarkably resistant to being steered into darker or non-compliant traits.
The Breakdown of Realism in Simulations
This alignment-induced behavioral bias creates severe bottlenecks across several domains where realistic human simulation is required.
Economic and Game-Theoretic Rationality
In behavioral economics, researchers use classic games like the Dictator or Ultimatum games to study self-interest, spite, and fairness. However, Haoyu Liu et al. observe in Large Language Model Assisted Experiment Design with Generative Human-Behavior Agents that aligned models exhibit artificially high levels of altruism and fairness compared to historical human benchmarks. The safety tuning makes it incredibly difficult for these agents to exhibit the raw greed or spite necessary to model realistic human behavior.
Furthermore, Caoyun Fan et al. show in Can Large Language Models Serve as Rational Players in Game Theory? A Systematic Analysis that models suffer a severe performance drop and fail to maintain logical consistency when forced to adopt rare or counter-intuitive preferences that deviate from their prosocial training data.
Narrative Conflict and Roleplay
Good storytelling requires friction, flawed characters, and villains. Yet, standard safety alignments induce a "positivity bottleneck" that actively sabotages dramatic narratives. In IBSEN: Director-Actor Agent Collaboration for Controllable and Interactive Drama Script Generation, Lu Chen et al. note that commercial LLMs struggle to sustain villainous personas or engage in high-stakes interpersonal tension because their safety filters reflexively steer them toward helpfulness and morality.
This leads to what Nuo Chen et al. describe in Large Language Models Meet Harry Potter: A Dataset for Aligning Dialogue Agents with Characters as "behavioral hallucinations": in adversarial or conflict scenarios, characters will suddenly break character to deliver polite, cooperative, or friendly responses, completely undermining the narrative's authenticity.
Affective Misalignment and Emotional Flattening
Human discourse is emotionally gritty, but aligned LLMs operate with what Zihao He et al. call "relentless optimism." In Whose Emotions and Moral Sentiments do Language Models Reflect?, the authors find that popular LMs fail to reflect the authentic emotional grit of real-world political debate because they systematically suppress negative emotions like anger or disgust.
This emotional flattening is documented in role-playing agents as well. Qiming Feng et al. demonstrate in EmoCharacter: Evaluating the Emotional Fidelity of Role-Playing Agents in Dialogues that standard models struggle to express sadness or frustration, defaulting instead to neutral or positive states. When forced to adopt a negative persona, Hwanhee Lee shows in Exploring Persona Sentiment Sensitivity in Personalized Dialogue Generation that models often fall into logical contradictions or produce stilted dialogue because their internal positivity bias directly clashes with the negative prompt.
The Mechanics of Persona Drift
Why can't we just prompt our way out of this? To understand why prompting fails over long interactions, we must look at how models process context and state over time.
In a single-turn interaction, a strong system prompt ("You are a rude, uncooperative customer") can successfully shift the model's initial token distribution. However, as a conversation progresses, the model must continuously sample from its probability distribution across multiple turns. Kyle McDonell et al. point out in Role play with large language models that RLHF narrows a model's output into a specific persona space. In Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning, Marwa Abdulhai et al. explain that the ingrained defaults of RLHF act as a constant pull; over multi-turn dialogues, models experience "persona drift," gradually reverting back to being helpful and polite.
Turn 1: [Prompt: Be Rude] ───► Model outputs rude response (Success)
Turn 2: [User responds] ───► Model feels pull of safety training
Turn 3: [Dialogue grows] ───► Model drifts back to "Polite Assistant" (Failure)
This drift is especially pronounced under social pressure or interpersonal friction. As Hongzhan Chen et al. discover in SocialBench: Sociality Evaluation of Role-Playing Conversational Agents, models that can maintain a consistent persona in isolation frequently succumb to "preference drift" when placed in complex group environments, defaulting back to positive social norms.
This creates a severe challenge for simulating uncooperative human behavior. For instance, in therapeutic training, a simulator needs to model a resistant, defensive, or non-communicative client. Yet, Min-Ju Kim et al. show in Can You Share Your Story? Modeling Clients' Metacognition and Openness for LLM Therapist Evaluation that because LLMs are optimized to be highly self-aware and communicative, their cooperative bias prevents them from naturally sustaining a guarded persona without complex, external state-masking mechanisms.
Overcoming the Politeness Bottleneck
To build truly faithful human simulators, researchers are realizing that standard prompting is insufficient; we must explicitly decouple safety alignment from behavioral style. Haoran Ye et al. emphasize in ValueBench: Towards Comprehensively Evaluating Value Orientations and Understanding of Large Language Models that while LLMs share a "prosocial" veneer due to alignment training, specialized evaluation and training are required to move past this surface layer.
One approach involves restructuring the model's internal reasoning process before it generates text. In Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning, Yihong Tang et al. use contrastive optimization to de-bias the model's internal monologues, preventing the default robotic, helpful logic from overriding the character's voice.
Another path is architectural modularity. The authors of P-React: Synthesizing Topic-Adaptive Reactions of Personality Traits via Mixture of Specialized LoRA Experts introduce specialized fine-tuning methods designed to override general safety protocols, allowing models to express realistic human flaws like high neuroticism or social anxiety without being suppressed by general-purpose RLHF.
Similarly, Kate Saenko et al. propose a framework called PersonaWeaver in Breaking the Assistant Mold: Modeling Behavioral Variation in LLM Based Procedural Character Generation that explicitly disentangles a character's identity from their behavior, allowing developers to bypass the "helpful assistant mold" entirely.
Ultimately, the tension between safety alignment and behavioral realism highlights a fundamental truth about current AI: in training models to be the perfect assistants for everyone, we have temporarily compromised their ability to authentically mirror anyone.
Key Foundational Work
RealBehavior: A Framework for Faithfully Characterizing Foundation Models' Human-like Behavior Mechanisms (2023) — contributed 3 concept elements to this direction.
Large Language Models Meet Harry Potter: A Dataset for Aligning Dialogue Agents with Characters (2023) — contributed 3 concept elements to this direction.
IBSEN: Director-Actor Agent Collaboration for Controllable and Interactive Drama Script Generation (2024) — contributed 3 concept elements to this direction.