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

Researchers systematically analyzed how leading LLMs (GPT-4o, Llama-3.3, Mistral-Large-2.1) generate demographically targeted messaging and found consistent gender and age-based biases, with male and youth-targeted messages emphasizing agency while female and senior-targeted messages stress tradition and care. The study demonstrates how demographic stereotypes intensify in realistic targeting scenarios, highlighting critical fairness concerns for AI-driven personalized communication.

This research addresses a fundamental blind spot in AI systems deployed at scale: how language models perpetuate demographic stereotypes when generating personalized content. While LLMs have been audited for various bias dimensions, the specific interaction between demographic conditioning and persuasive messaging generation reveals a compounding problem. When tasked with tailoring climate communication messages, all three tested models exhibited systematic asymmetries that align with historical gender and age stereotypes rather than demonstrating neutral, evidence-based adaptation.

The distinction between Standalone Generation and Context-Rich Generation is particularly revealing. Adding thematic and regional context amplified rather than mitigated bias, suggesting that realistic deployment scenarios exacerbate stereotype surfacing. This pattern indicates the problem isn't isolated incidents but structural: the models internalized demographic associations from training data and activate them reliably when prompted to personalize. The persuasion score differential favoring younger and male audiences compounds the fairness problem, as marketers and communicators using these tools would systematically achieve better outcomes targeting privileged demographics.

For developers and enterprises deploying personalization at scale—from marketing to political messaging to public health campaigns—this research creates immediate compliance and reputational risks. Regulatory frameworks around algorithmic fairness are tightening globally, and organizations using biased targeting systems face litigation exposure and brand damage. The paper's call for bias-aware generation pipelines and transparent auditing frameworks suggests the market will demand fairness-focused alternatives. However, the current lack of standardized mitigation strategies creates both a challenge for deployment and an opportunity for tool developers who can offer genuinely debiased personalization solutions.

• →All three leading LLMs consistently generated gender and age-based stereotypical messages, indicating widespread rather than isolated bias
• →Contextual prompts systematically amplified demographic disparities, meaning real-world deployment scenarios intensify rather than reduce bias
• →Male and younger audiences received messages with higher persuasion scores, creating fairness risks in marketing and political communication
• →Current LLM architectures lack built-in safeguards against stereotype activation during demographic conditioning tasks
• →Organizations deploying personalized AI messaging face growing regulatory and reputational risks without explicit bias auditing frameworks