Political Bias in Large Language Models: A Study on AI Bias from the Perspective of India's Gen Z

This article focuses on the study of political bias in large language models from the perspective of India's Gen Z. It corely explores the systemic biased performance of mainstream large language models in Indian political narratives, their impact on Gen Z's political cognition, the root causes of bias, and mitigation strategies. It aims to reveal the potential harm of AI bias to democratic participation and information dissemination, and provide references for AI governance.

With the widespread application of large language models globally, the issue of political bias in AI systems has received increasing attention. As the world's largest democracy, India has a complex political ecosystem and diverse social structure, providing a unique case for studying AI political bias; as digital natives, Gen Z's political views are closely related to AI technology, making them ideal subjects for studying the impact of AI bias.

This study adopts a mixed-method approach: 1. Model bias detection: Select mainstream large language models, design prompt templates covering India's major political parties, policy issues, and political figures, and analyze responses to neutral prompts to identify stance tendencies; 2. Gen Z cognition survey: Conduct large-scale questionnaires and in-depth interviews with India's Gen Z to understand the frequency, trust, and impact of their use of AI to obtain political information; 3. Data source analysis: Trace the sources of model training data to identify the root causes of bias.

Model-level Bias

• Party tendency: Some models use more positive/negative language for specific parties;
• Issue framing: Tend to frame political issues in specific ways;
• Selective historical narration: Information bias exists when involving historical events

Impact on Gen Z's Cognition

• Over 70% of respondents often use AI to obtain political information;
• Less than 30% are aware that AI may have political bias;
• Exposure to biased content is significantly correlated with political opinion polarization

Data Root Causes

• English data dominance leads to deviations in understanding local context;
• Western media reports bring external perspective bias;
• Unbalanced historical data leads to distortion in understanding contemporary politics

Data Level

• Increase training data of local multilingual political texts in India;
• Balance data sources of different political stances;
• Introduce fact-checking mechanisms to filter misinformation

Model Level

• Develop politically neutral fine-tuning technology;
• Implement bias detection and early warning systems;
• Introduce multi-perspective generation mechanisms

User Level

• Label bias risks on AI interfaces;
• Provide information source tracing functions;
• Carry out digital literacy education to improve critical evaluation ability

Necessity of Regulatory Framework

Need to establish bias assessment standards and independent audit mechanisms, and require AI service providers to disclose potential bias risks

Importance of Multicultural Perspectives

Development teams with a single cultural background find it difficult to identify biases in other cultures; thus, multiple perspectives need to be included

Protection of Democratic Participation

AI bias may erode citizens' right to access unbiased political information and needs to be included in the agenda of digital rights and democratic governance

Limitations

• Samples are concentrated on urban Gen Z, with insufficient representation of rural and marginalized groups;
• The research time point is specific, making it difficult to capture long-term trends;
• Technical methods are based on existing models; new architectures may change the performance of bias

Future Directions

Expand to other countries/regions to explore the impact of cultural differences; develop more effective bias detection and mitigation technologies