Are Large Language Models Really Bias-Free? Jailbreak Prompts for Assessing Adversarial Robustness to Bias Elicitation

2024-07-11

IFIP Working Conference on Database Semantics

AI Summary

This work systematically investigates the elicitation of implicit social biases in large language models (LLMs). Addressing the limitation of prior research—which predominantly examines explicit biases and lacks controllable elicitation mechanisms—we propose the first taxonomy and evaluation benchmark specifically designed for bias elicitation. Our method introduces a multi-strategy prompt engineering framework integrating semantic perturbation, role-playing, and context injection, evaluated via dual-track assessment combining human annotation and inter-annotator consistency checks. Experiments across major open- and closed-source LLMs demonstrate that our approach increases bias response rates by an average factor of 3.8×, effectively exposing critical robustness deficiencies beneath models’ superficial neutrality. This work establishes a reproducible methodological foundation and empirical evidence for bias detection, fairness evaluation, and alignment optimization in LLMs.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Philosophy and Ethics of AI: Bias, Fairness & Equity

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Problem

• Assessing biases in Large Language Models
• Testing adversarial robustness with jailbreak prompts
• Enhancing mitigation techniques for bias reduction

Innovation

• Jailbreak prompts reveal biases
• Analyze LLM adversarial robustness
• Enhance bias mitigation techniques