This research introduces a framework to reduce discriminatory responses in large language models by identifying and removing causal relationships between social information and model decisions.

Key Findings:

• Leverages causal inference to identify and mitigate biases in LLM responses
• Targets critical applications in high-stakes decision-making including hiring and healthcare
• Reduces objectionable dependencies between social information inputs and model outputs
• Demonstrates a novel approach to fairness that can be implemented through effective prompting

Security Implications:
As LLMs increasingly influence important decisions, this framework addresses a critical security concern by preventing discriminatory outputs that could lead to unfair treatment or legal liability in sensitive domains.

Prompting Fairness: Integrating Causality to Debias Large Language Models