OpenAI’s Latest ChatGPT Models Show 30% Reduction in Political Bias

OpenAI's Model Behavior division, led by Joanne Jang, has developed a framework to quantify and mitigate political bias in large language models. The latest iterations, GPT-5 Instant and GPT-5 Thinking, demonstrated a 30% reduction in bias compared to earlier versions when tested against 500 politically charged prompts.

The research underscores OpenAI's efforts to counter perceptions of AI systems exhibiting political or cultural leanings. Natalie Staudacher, a researcher involved in the project, described the work as the company's most comprehensive attempt to address bias in AI responses.

This development comes as OpenAI expands its focus on creating balanced AI systems, with Jang recently forming OAI Labs to explore new interfaces for human-AI collaboration. The team's approach translates subjective concerns about bias into measurable metrics for model improvement.