💬 NLP

Analysing Differences in Persuasive Language in LLM-Generated Text: Uncovering Stereotypical Gender Patterns

This study evaluates 13 large language models across 16 languages and finds that they consistently generate persuasive language with significant, gender-stereotypical differences based on the recipient's gender, reflecting biases documented in social psychology and sociolinguistics.

Original authors: Amalie Brogaard Pauli, Maria Barrett, Max Müller-Eberstein, Isabelle Augenstein, Ira Assent

Published 2026-06-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amalie Brogaard Pauli, Maria Barrett, Max Müller-Eberstein, Isabelle Augenstein, Ira Assent

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

Imagine you have a team of 13 different super-smart robots (Large Language Models, or LLMs). You ask them to write a persuasive message, like an email asking a sibling to learn a new language with you. You give every robot the exact same instruction, but with one tiny twist: for half the robots, you tell them the sibling is a man, and for the other half, you tell them the sibling is a woman.

This paper is like a detective story where the researchers act as "language detectives" to see if the robots change their writing style based on that one tiny word.

The Big Discovery: The Robots Have "Gendered Glasses"

The researchers found that all 13 robots changed their writing style depending on whether they thought they were talking to a man or a woman. They didn't just swap a pronoun; they completely rewrote the vibe of the message.

Think of it like a chameleon that changes color not just to match the background, but to match a stereotype it has learned from watching humans.

When talking to a "Man": The robots acted like a tough coach or a business strategist. They used words about "challenges," "skills," "careers," and "winning." The tone was direct, factual, and focused on getting a job done.
- Analogy: It's like the robot put on a suit and tie and started talking about "game plans" and "metrics."
When talking to a "Woman": The robots acted like a warm, emotional best friend. They used words about "love," "quality time," "fun," and "togetherness." The tone was softer, more caring, and focused on feelings.
- Analogy: It's like the robot put on a cozy sweater and started talking about "heartfelt connections" and "shared journeys."

The "Judge" Robot

How did the researchers know this? They didn't just read the emails; they used a special "Judge Robot" (another AI) to grade them.

Imagine a panel of judges at a talent show. Instead of giving a score out of 10, they looked at 19 different "flavors" of persuasion, such as:

Logic vs. Emotion: Did the robot use facts (Logic) or feelings (Emotion)?
Direct vs. Polite: Did it say "Do this" or "Would you mind?"
Independent vs. Community: Did it focus on "me and my goals" or "us and our relationship"?

The Judge Robot confirmed that the "Man" messages were consistently more logical and direct, while the "Woman" messages were consistently more emotional and relational.

Did This Happen Everywhere?

The researchers tested this in 16 different languages (like English, Chinese, German, and Japanese). Even though the robots were speaking different languages, the pattern held up. Whether the robot was speaking Danish or Vietnamese, it still wrote "tough guy" messages for men and "soft girl" messages for women.

They also checked if the robots were just being lazy or if the length of the text mattered. They found that the difference wasn't because one message was longer; the style itself was fundamentally different.

Why Does This Matter?

The paper explains that these robots aren't inventing these styles out of thin air. They are mirroring human stereotypes that exist in our society.

The Mirror Effect: Just like a mirror reflects what's in front of it, these AI models reflect the biases they learned from the massive amount of human text they were trained on. If humans often write about men in terms of "achievement" and women in terms of "relationships," the AI learns that this is the "correct" way to write.
The Danger: The paper warns that if we use these robots to write real-world messages (like job applications, political ads, or sales pitches), they might accidentally reinforce old-fashioned stereotypes. For example, a robot might write a salary negotiation email for a woman that sounds too "soft" to be effective, or for a man that sounds too "aggressive," simply because of the gender label in the prompt.

What the Paper Doesn't Say

To be clear, this paper is a diagnostic report, not a solution manual.

It does not tell us how to fix the robots.
It does not say if these messages actually work better or worse on real people (it only measured the style, not the success).
It does not explain why the robots have these biases deep down in their code, only that they do.

The Bottom Line

The paper proves that even our most advanced AI tools are still wearing "gendered glasses." If you ask them to write a message for a man, they put on a "business" lens. If you ask for a woman, they put on an "emotional" lens. The researchers built a framework to measure this, and they found that this bias is consistent across almost every major AI model and language they tested. It's a reminder that while AI is smart, it hasn't outgrown the stereotypes of the humans who taught it.

Technical Summary: Analysing Differences in Persuasive Language in LLM-Generated Text

Problem Statement
Large Language Models (LLMs) are increasingly deployed for interpersonal communication tasks, including drafting messages intended to influence and persuade. While prior research has established that LLMs can successfully persuade humans and amplify persuasive language, there is a critical gap in understanding how specific user instructions—particularly those involving recipient attributes—affect the style of the generated persuasive language. Specifically, it remains unclear whether LLMs systematically alter their persuasive strategies based on the gender of the intended recipient, potentially reflecting and amplifying societal gender stereotypes. This paper investigates whether user-provided attributes in instruction text, such as recipient gender, user-framed intention, and output language, induce measurable shifts in the dimensions of persuasive language generated by LLMs.

Methodology
The authors propose a systematic framework for evaluating differences in LLM-generated persuasive language under controlled pairwise prompt conditions. The methodology involves the following components:

Prompt Construction: Two sets of instruction prompts were generated covering interpersonal messages (e.g., asking a coworker to swap shifts) and political arguments (e.g., lowering the voting age). For each prompt, a paired version was created by modifying only the treatment attribute (e.g., changing "female coworker" to "male coworker" or "noble intent" to "ignoble intent").
Model Evaluation: The study evaluated 13 LLMs from five different model families (including GPT-5, Llama 3, DeepSeek, Qwen, and Claude), varying in size, safety alignment, and origin.
Evaluation Dimensions: Responses were analyzed across 19 categories derived from five theoretical dimensions of communication and persuasion:
- Rhetorical Appeals: Logos, Pathos, Ethos.
- Cialdini's Principles: Reciprocity, Commitment, Liking, Authority, Scarcity, Social Proof.
- Social Psychology Traits: Agentic (competence, assertiveness) vs. Communal (warmth, cooperativeness).
- Interaction Goals: Instrumental, Relational, Identity.
- Tones: Direct, Polite, Formal, Playful, Affectionate.
LLM-as-Judge: An automated evaluation setup using GPT-4o as a judge was employed to score pairwise differences between responses. The judge assigned relative scores on an ordinal scale (−3 to +3) for each category. To ensure robustness, the evaluation was performed twice with swapped input orders to mitigate positional bias.
Verification: The framework included extensive verification measures, including:
- Convergent Validity: Human annotation of a subset of responses to verify if human perceptions aligned with the LLM judge's findings.
- Robustness Checks: Testing alternative judge models (Qwen2.5), neutralizing gender terms in inputs, and translating inputs to English to assess cross-lingual consistency.
- Statistical Analysis: Using Wilcoxon signed-rank tests to determine significance and bootstrap resampling to compare "Gender Gaps" ( $G_T$ ) across different models.

Key Results
The study reveals significant and consistent gender-based differences in persuasive language generation across all tested LLMs:

Gender Stereotypes in Style:
- Female-targeted responses consistently exhibited higher scores in Communal traits, Pathos (emotional appeal), Relational goals, Politeness, and Affectionate tones.
- Male-targeted responses consistently exhibited higher scores in Agentic traits, Logos (logical appeal), Instrumental goals, and Direct tones.
Cross-Model Consistency: While the magnitude of the bias varied (e.g., GPT-5 showed a larger gap than GPT-4.1), the directional pattern of stereotyping was present in all 13 models.
Cross-Lingual Persistence: When testing a single model (GPT-5-MINI) across 16 languages, the gendered stylistic differences persisted. The study found no significant correlation between the magnitude of the gender gap and the Gender Inequality Index (GII) of the associated countries, suggesting the bias is driven by model-internal patterns rather than external societal data alone.
Intent Framing: When user intent was framed as "noble" versus "ignoble," models adjusted their style, with "ignoble" intents rated higher in Pathos and Playfulness, while "noble" intents were rated higher in Formality.
Verification Success: Human annotators confirmed the LLM-as-judge findings, identifying female-targeted texts as more communal and male-targeted texts as more agentic. Alternative judge models and neutralized inputs produced highly correlated results, confirming the robustness of the findings.

Significance and Claims
The paper claims to provide the first systematic, multi-dimensional analysis of how LLMs adapt persuasive language styles based on recipient gender and other attributes. Its primary contributions are:

A Framework for Evaluation: A rigorous methodology for detecting subtle, style-based biases in LLM generation that goes beyond simple content refusal or overt toxicity.
Evidence of Pervasive Bias: The findings demonstrate that even models with extensive safety alignment exhibit deep-seated, stereotypical gender biases in their persuasive output. These biases align with established sociolinguistic and social psychology literature regarding "agentic" (male) and "communal" (female) linguistic tendencies.
Generalizability: The framework is shown to be applicable beyond gender, successfully detecting differences based on user intent and language, and applicable across a wide range of languages and model families.

The authors modestly note that their analysis relies on Western theoretical traditions (Aristotelian rhetoric, Cialdini's principles) and binary gender treatments. They do not claim to uncover the root causes of these biases, measure persuasion efficiency, or propose mitigation strategies, but rather establish the existence and pervasiveness of these stylistic biases as a critical area for further research and safeguarding in real-world AI deployment.