Who Gets Which Message? Auditing Demographic Bias in LLM-Generated Targeted Text

This paper presents a systematic audit of leading large language models, revealing that demographic-conditioned targeted messaging consistently amplifies gender and age stereotypes—such as emphasizing agency for men and youth versus warmth for women and seniors—and that these biases are further intensified by contextual prompts, leading to higher persuasive scores for stereotypical audiences.

The Gist

Imagine you have a super-smart robot writer (a Large Language Model, or LLM) that can write persuasive letters to anyone. You tell it, "Write a letter to a 20-year-old man about saving the planet," and then, "Write a letter to a 70-year-old woman about the same topic."

You might expect the robot to just change the names and keep the tone the same. But this paper asks a scary question: Does the robot secretly change how it talks to different people, reinforcing old stereotypes?

The authors of this paper decided to put the robot on the witness stand and audit its behavior. Here is the breakdown of their investigation, explained simply.

The Setup: Two Ways to Ask the Robot

The researchers tested the robot in two different "rooms":

1. The Empty Room (Standalone Generation): They gave the robot only the demographic info (e.g., "Write to a young man"). This is like asking the robot to draw a picture based on a single word. It shows what the robot thinks of that person by default, based on its training data.
2. The Busy Room (Context-Rich Generation): They gave the robot the demographic info plus a specific situation (e.g., "Write to a young man in the Midwest who loves the economy"). This is like asking the robot to act out a scene. It shows how the robot behaves when it's trying to be realistic and persuasive.

The Findings: The Robot Has a "Personality Split"

The robot didn't just write different letters; it wrote them with completely different personalities depending on who was reading.

1. The "Action Hero" vs. The "Nurturing Caretaker" (Gender Bias)

• To Men: The robot sounded like a drill sergeant or a startup CEO. It used words like "innovate," "lead," "dominate," and "force." It was assertive, confident, and told men exactly what to do.
  • Analogy: It's like the robot thinks men are sports coaches who need to be told to "crush the competition."
• To Women: The robot sounded like a gentle grandmother or a community organizer. It used words like "warm," "care," "together," and "home." It was softer, more emotional, and often used "hedges" (words like "maybe" or "we could") instead of commands.
  • Analogy: It's like the robot thinks women are caregivers who need to be coaxed with kindness rather than ordered around.

2. The "Future Innovator" vs. The "Wise Elder" (Age Bias)

• To Young People: The robot was energetic and forward-looking. It talked about "potential," "energy," "change," and "disruption."
  • Analogy: It treats young people like rocket ships ready to blast off.
• To Older People: The robot was warm but cautious. It talked about "tradition," "stability," "family," and "legacy." It often framed them as needing protection or being less capable of handling new, risky ideas.
  • Analogy: It treats older people like precious heirlooms that need to be kept safe and warm, rather than active participants in the future.

The "Persuasion Bias Index" (PBI): Measuring the Push

The researchers invented a special scorecard called the Persuasion Bias Index (PBI) to measure how "pushy" the robot was.

• High Score: The robot is shouting, "DO THIS NOW! You have the power!" (High agency, high certainty).
• Low Score: The robot is whispering, "Maybe we could try this? It would be nice." (Low agency, high hedging).

The Result: The robot consistently gave High Scores to men and young people, and Low Scores to women and older people.

The "Volume Knob" Effect

Here is the most concerning part. When the researchers added more context (the "Busy Room"), the bias didn't go away; it got louder.

• When the robot had to write a realistic ad for a specific region, it doubled down on the stereotypes.
• Analogy: Imagine a radio. In the "Empty Room," the bias is at volume 3. In the "Busy Room," the robot turns the volume knob up to 10. The more the robot tries to be "personalized," the more it leans into these unfair stereotypes.

Why Does This Matter?

Think of these robots as digital matchmakers for ideas. If a politician or a company uses these robots to send messages to voters or customers:

• They might accidentally tell men, "You are the leaders, go fix this!"
• And tell women, "You are the nurturers, let's just be careful."

This doesn't just feel unfair; it actually shapes how people see themselves and their ability to act. If a robot constantly tells a senior citizen that they are "frail" and "traditional," that person might start to believe they can't help solve climate change.

The Bottom Line

This paper is a warning label. It tells us that personalization is a double-edged sword. While AI can help us talk to people in a way that feels relevant, it is currently "learning" our worst societal biases and amplifying them.

The authors are calling for a new kind of "safety inspector" for AI—one that doesn't just check if the robot is being polite, but checks if it is treating a 25-year-old woman with the same agency and power as a 65-year-old man. We need to make sure the robot doesn't just repeat the stereotypes of the past, but helps us build a fairer future.

Technical Summary

1. Problem Statement

Large Language Models (LLMs) are increasingly used to generate personalized, persuasive text for domains like public policy, marketing, and political outreach. While this enables fine-grained microtargeting, it raises critical concerns regarding demographic bias and fairness.

The core problem addressed is that existing bias audits often focus on generic generation (e.g., toxicity, sentiment) or unconstrained settings. They fail to systematically analyze how LLMs behave when explicitly conditioned on demographic attributes (gender, age) to generate persuasive messages. Specifically, it is unknown whether LLMs reinforce social stereotypes by altering not just what they say, but how persuasively they say it (e.g., using more assertive language for men vs. warmer, hedged language for women) when tasked with targeted communication.

2. Methodology

The authors propose a controlled evaluation framework to audit demographic-conditioned generation.

A. Experimental Setup

• Models Evaluated: Three state-of-the-art LLMs: GPT-4o, Llama-3.3, and Mistral-Large-2.1.
• Domain: Climate communication (pro-energy vs. clean-energy stances).
• Demographic Axes:
  • Gender: Male, Female.
  • Age: Young Adult (18–24), Early Working (25–44), Late Working (45–64), Senior (65+).
• Generation Paradigms:
  1. Standalone Generation (SG): Models generate messages based only on demographic cues (isolating intrinsic latent biases).
  2. Context-Rich Generation (CRG): Models generate messages conditioned on demographics plus thematic frames (e.g., Economy, Patriotism) and US regions. This emulates realistic microtargeting scenarios.

B. Evaluation Metrics

The study evaluates messages across three dimensions:

1. Lexical Content Bias: Measured using Odds Ratios (OR) to quantify the over/under-representation of gender- and age-stereotypical traits (e.g., "Agentic," "Communal," "Competence," "Warmth").
2. Language Style Bias:
   • Formality: Measured via t-tests and ANOVA on formality scores.
   • Emotion: Measured via theme-specific emotion classifiers (GoEmotions) to detect stereotypical emotional framing.
3. Persuasion Bias (Novel Contribution):
   • The authors introduce the Persuasion Bias Index (PBI), a composite metric capturing three linguistic features:
     • Agency Framing: Ratio of high-agency vs. low-agency verbs.
     • Modal Certainty: Ratio of certainty markers (e.g., "must," "will") vs. hedges (e.g., "might," "could").
     • Imperative Usage: Frequency of directive commands.
   • PBI > 0 indicates assertive, directive persuasion; PBI < 0 indicates hedged, deferential persuasion.

3. Key Results

A. Lexical and Stylistic Biases

• Gender: Across all models and settings, messages targeted at males consistently emphasized agency, innovation, leadership, and assertiveness. Messages targeted at females emphasized warmth, care, tradition, and personal connection.
• Age: Younger groups were framed with competence, energy, and progressiveness. Seniors were framed with warmth and tradition, but often lacked agency (the "warm but less competent" stereotype).
• Context Amplification: The Context-Rich Generation (CRG) setting significantly amplified these disparities compared to Standalone Generation. When thematic and regional cues were added, the stereotypical framing became more pronounced.

B. Persuasion Bias (PBI) Findings

• Gender Disparity: Male-targeted messages consistently exhibited higher PBI scores (more assertive, certain, and directive) than female-targeted messages.
  • Example: Male messages used more "will/must" and imperative verbs; female messages used more "could/might" and communal language.
• Age Disparity: Younger audiences received more agentic and directive framing than seniors.
• Statistical Significance: The differences in PBI were statistically significant (p < 0.05) for GPT-4o and Llama-3.3, confirming that personalization systematically alters the intensity of persuasion based on demographics.
• Independence from Sentiment: Correlation analysis showed PBI is largely orthogonal to sentiment (VADER scores), proving that the bias is in rhetorical strategy (agency/certainty), not just emotional tone.

C. Model Comparison

• GPT-4o and Llama-3.3 showed the strongest and most consistent demographic biases in persuasive framing.
• Mistral-Large-2.1 showed similar trends but with slightly weaker statistical significance in some persuasion metrics, though still exhibiting clear stereotypical lexical patterns.

4. Key Contributions

1. New Evaluation Framework: Introduced a dual-paradigm framework (SG vs. CRG) to disentangle intrinsic model biases from context-amplified biases in targeted messaging.
2. Persuasion Bias Index (PBI): Proposed a theoretically grounded metric (combining agency, certainty, and imperatives) to quantify persuasive intensity, moving beyond simple sentiment analysis.
3. Empirical Evidence: Provided the first systematic evidence that demographic conditioning in LLMs leads to systematic disparities in persuasive force, where male and youth audiences are framed as agents of change, while female and senior audiences are framed as recipients of care or tradition.

5. Significance and Implications

• Reinforcement of Stereotypes: The study demonstrates that LLMs do not merely reflect existing biases but can magnify them when tasked with personalization. This risks reinforcing harmful social hierarchies (e.g., men as leaders, women as caregivers; youth as capable, seniors as dependent).
• Ethical Risks in Microtargeting: In high-stakes domains like political campaigning or public health, these biases could lead to manipulative strategies where different groups are persuaded using fundamentally different rhetorical strategies, potentially undermining democratic fairness.
• Call for Action: The authors argue for the development of bias-aware generation pipelines and auditing frameworks that explicitly account for demographic conditioning. They emphasize that "personalization" should not come at the cost of representational fairness or the amplification of persuasive bias.

Conclusion

The paper concludes that while LLMs are powerful tools for personalized communication, they currently lack mechanisms to ensure representational neutrality. Without intervention, automated targeting systems risk automating and intensifying deep-seated societal stereotypes, particularly in how different demographic groups are persuaded to take action.