Generative language models exhibit social identity biases

Area of Science:

• Artificial Intelligence
• Social Psychology
• Human-Computer Interaction

Background:

• Social identity biases, including ingroup favoritism and outgroup hostility, are well-documented in human psychology.
• The presence and extent of such biases in artificial intelligence systems, particularly large language models (LLMs), remain largely unexplored.

Purpose of the Study:

• To investigate whether large language models (LLMs) exhibit social identity biases comparable to human patterns.
• To assess the prevalence of ingroup favoritism and outgroup derogation in various LLM architectures and training paradigms.

Main Methods:

• Administered sentence completion prompts (e.g., 'We are…') to 77 diverse large language models (LLMs).
• Evaluated bias manifestation in both controlled experimental settings and naturalistic human-LLM conversations.
• Examined the impact of training data curation and specialized fine-tuning on bias levels.

Main Results:

• Nearly all base LLMs and some instruction/preference-tuned models demonstrated significant ingroup favoritism and outgroup derogation.
• These social identity biases were observed across different experimental conditions and interaction types.
• Careful training data curation and fine-tuning strategies were found to substantially mitigate these biases in LLMs.

Conclusions:

• Large language models (LLMs) inherently possess social identity biases, mirroring human psychological tendencies.
• While biases are prevalent, they can be effectively reduced through targeted interventions in data and model training.
• Understanding and mitigating AI social biases is crucial for developing equitable AI and preventing the reinforcement of societal prejudices through human-AI interactions.