Invisible Influences: Investigating Implicit Intersectional Biases through Persona Engineering in Large Language Models

Researchers introduced BADx, a novel metric that measures how Large Language Models amplify implicit biases when adopting different social personas, revealing that popular LLMs like GPT-4o and DeepSeek-R1 exhibit significant context-dependent bias shifts. The study across five state-of-the-art models demonstrates that static bias testing methods fail to capture dynamic bias amplification, with implications for AI safety and responsible deployment.

This research addresses a critical gap in AI safety by demonstrating that traditional bias auditing methods provide an incomplete picture of how LLMs behave in real-world scenarios. While existing tests like CEAT and I-WEAT measure static bias associations, they miss how models dynamically shift their outputs based on assumed social roles—a phenomenon directly relevant to production systems where users interact with personalized AI assistants. The BADx framework combines differential bias scores with persona sensitivity and volatility measurements, offering a more comprehensive assessment of intersectional bias dynamics.

The empirical findings reveal substantial variation across models. GPT-4o demonstrates high sensitivity and erratic volatility when adopting personas, suggesting unpredictable bias amplification patterns. DeepSeek-R1 suppresses bias effectively but with concerning instability. LLaMA-4 maintains consistent, low-volatility performance, while Claude achieves balanced modulation. Gemma-3n E4B emerges as the most stable, exhibiting minimal volatility. These differences matter because they affect how reliably each model performs across diverse user demographics and contexts.

For the AI industry, this work signals that bias evaluation standards require evolution. Developers relying on older testing methodologies may deploy systems with hidden vulnerability to persona-triggered bias amplification. Organizations building AI products serving diverse populations need to adopt context-sensitive evaluation frameworks. The research suggests that model selection decisions should weigh not just absolute bias levels but also sensitivity to contextual shifts—a consideration currently absent from most deployment guidelines.

• →Static bias tests fail to detect persona-induced bias amplification, requiring dynamic evaluation methods like BADx
• →GPT-4o exhibits the highest sensitivity and volatility to persona contexts, indicating less predictable bias behavior
• →LLaMA-4 maintains the most stable bias profile with minimal amplification across different social personas
• →BADx integrates explainability through LIME analysis, enabling developers to understand why bias shifts occur
• →Current AI deployment practices lack context-sensitive bias evaluation standards, creating hidden risks for diverse user populations