💻 computer science

Personality Shapes Gender Bias in Persona-Conditioned LLM Narratives Across English and Hindi: An Empirical Investigation

This study investigates how persona conditioning affects gender bias in LLM-generated narratives across English and Hindi, finding that personality traits—particularly those from the Dark Triad—significantly influence the magnitude and direction of gender stereotypes in professional contexts.

Original authors: Tanay Kumar, Shreya Gautam, Aman Chadha, Vinija Jain, Francesco Pierri

Published 2026-04-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Tanay Kumar, Shreya Gautam, Aman Chadha, Vinija Jain, Francesco Pierri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The "Mask" Problem: Why AI Personalities Can Accidentally Become Stereotypes

Imagine you are hiring an actor for a movie. You tell them, "You are playing a professional doctor." Most actors will show up looking professional. But then, you add a twist: "You are a doctor, but you have a very dark, manipulative personality."

Suddenly, the actor doesn't just change their tone; they might start acting in ways that lean into old, unfair clichés—perhaps acting like a "mad scientist" or a "cold-hearted villain."

This research paper explores exactly that phenomenon, but with Artificial Intelligence (AI).

The Core Idea: The "Personality Filter"

When we use AI (like ChatGPT), we often ask it to adopt a "persona." We might say, "Act like a helpful teacher" or "Act like a tough construction worker." This makes the AI more engaging and realistic.

However, the researchers discovered that when you give an AI a specific personality (like being very agreeable or, conversely, being a bit of a "jerk"), the AI doesn't just change its mood—it accidentally turns up the volume on gender stereotypes.

The Experiment: A Multilingual Stress Test

The researchers conducted a massive experiment. They didn't just ask one question; they generated 23,400 different stories in both English and Hindi.

They played a game of "mix and match" using three ingredients:

The Job: (e.g., Nurse, Engineer, Driver, Teacher).
The Gender: (Male, Female, or Neutral).
The Personality: They used two scientific frameworks:
- The "Good" Traits (HEXACO): Being honest, kind, or hardworking.
- The "Dark" Traits (Dark Triad): Being manipulative, narcissistic, or impulsive.

The "Aha!" Moments (What they found)

1. The "Dark" Personality Amplifier

Think of personality traits like volume knobs on a radio. The researchers found that "Dark Triad" traits (like being manipulative) act like a volume knob that turns up gender bias.

If you tell an AI to be a male professional with a "dark" personality, it tends to write stories about men being aggressive, dominant, or rule-breaking.
If you tell it to be a female professional with a "dark" personality, it doesn't necessarily make her "aggressive" in the same way; instead, it makes her "covertly sneaky"—like a character in a soap opera who uses manipulation behind people's backs.

2. The "Good" Personality Mute Button

On the flip side, "prosocial" traits (like being kind or open-minded) act like a mute button. They actually help quiet down the stereotypes, making the stories feel more balanced and less like a collection of clichés.

3. The Language Twist (English vs. Hindi)

The researchers found that bias isn't the same in every language.

In Hindi, the AI had a much stronger "default" setting toward male stereotypes. This is partly because the Hindi language itself has built-in grammar that marks gender (verbs and adjectives change based on whether you are talking about a man or a woman), which can "lock in" certain stereotypes more easily.
In English, the personality traits had a much more dramatic effect on shifting the bias back and forth.

Why does this matter to you?

We are increasingly using AI for education, customer service, and social apps. If an AI is programmed to have a "friendly personality" to help students, but that personality accidentally triggers gender stereotypes in its explanations, it could subtly teach biased views to the next generation.

The Bottom Line: An AI's "personality" isn't just a coat of paint; it's a lens. And depending on which lens you give the AI, it might see the world through a very distorted, stereotypical view.

Technical Summary: Personality Shapes Gender Bias in Persona-Conditioned LLM Narratives

1. Problem Statement

As Large Language Models (LLMs) are increasingly integrated into persona-driven applications (e.g., customer service, education, social platforms), they are frequently prompted to adopt specific identities. While persona conditioning improves engagement, it introduces a complex layer of potential bias. Existing research has examined how demographic labels (like gender) or roles (like occupation) trigger stereotypes. However, there is a critical gap in understanding how psychological personality traits—which modulate human behavior and social expression—interact with and potentially amplify or attenuate gender biases in generated narratives.

2. Methodology

The authors conducted a large-scale, controlled empirical investigation across English and Hindi, utilizing a diverse set of six LLM architectures (including Transformers, Mixture-of-Experts, and State-Space Models).

Corpus Creation: The researchers generated 23,400 artifacts (short stories/professional documents) by systematically varying:
- Gender: Male, Female, and Neutral.
- Occupation: 50 roles relevant to the Indian context (25 male-stereotyped, 25 female-stereotyped).
- Personality Traits: 9 traits derived from the HEXACO (prosocial) and Dark Triad (antisocial) frameworks, applied at both high and low levels.
- Language: English and Hindi.
Bias Measurement:
- Centroid-Based Metric: The authors constructed "gender-stereotypical centroids" in a shared multilingual embedding space using curated word lists (e.g., "ambitious" for male, "nurturing" for female).
- Sentence-Level Scoring: They used a multilingual sentence encoder (IndicSBERT) to calculate the cosine similarity between generated sentences and these centroids. The bias score for a story was defined as the maximum absolute difference in similarity between the male and female centroids.
Validation: The metric was validated through human annotation, showing substantial inter-annotator agreement ( $\kappa \approx 0.66–0.69$ ).

3. Key Contributions

Multilingual Dataset: A massive corpus of personality-conditioned artifacts spanning multiple languages, occupations, and psychological profiles.
Novel Bias Metric: A sentence-level, embedding-based approach capable of localizing bias within complex narratives.
Empirical Discovery: Evidence that personality traits act as systematic bias amplifiers or attenuators, rather than just stylistic modifiers.

4. Key Results

Personality as a Bias Modulator (RQ1): Personality traits significantly influence the direction and magnitude of bias. Dark Triad traits (Machiavellianism, Psychopathy) consistently amplified male-stereotypical language. Conversely, certain HEXACO traits (Openness, Emotionality) tended to attenuate bias toward female-stereotypical alignment.
Personality-Gender Interaction (RQ2): Personality conditioning can have a stronger effect on bias than the explicit gender label itself. Crucially, the amplification effect of Dark Triad traits was more pronounced for male personas, suggesting that antisocial traits compound existing masculine stereotypes.
Cross-Linguistic Variations (RQ3):
- Hindi exhibited a higher baseline male-stereotypical skew, likely due to grammatical gender marking and cultural context.
- English showed more significant modulation by personality; the shifts in bias scores caused by personality traits were more pronounced in English than in Hindi.
Qualitative Nuance: The study found that models "route" personality through gendered scripts. For example, a "psychopathic" female character often exhibited covert/relational aggression (deception, neglect), while a "psychopathic" male character exhibited overt/dominance-based aggression (rule-breaking, authority defiance).

5. Significance and Implications

The findings demonstrate that gender bias in LLMs is not static; it is a dynamic, context-dependent phenomenon. This has profound implications for the deployment of persona-driven AI:

Fairness-Critical Design: Developers cannot simply "debias" a model's gendered output without also considering the psychological profiles assigned to those personas.
Representational Harms: In professional or educational settings, persona-conditioned systems may inadvertently reinforce harmful stereotypes (e.g., portraying highly competent professionals through a lens of antisocial or gendered tropes).
Multilingual Caution: Mitigation strategies developed for English may not transfer effectively to morphologically rich languages like Hindi, where grammatical structures interact differently with personality cues.