OpenAI has detailed its approach to defining and evaluating political bias in its large language models, aiming for objectivity by default with user control. The company developed a new evaluation framework using approximately 500 prompts across 100 topics to measure five nuanced axes of bias. Initial results show that while models remain near-objective on neutral prompts, they exhibit moderate bias on emotionally charged ones, with GPT-5 variants demonstrating a 30% reduction in bias compared to prior models. OpenAI estimates that less than 0.01% of real-world ChatGPT responses show signs of political bias and plans further improvements, particularly for challenging prompts. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
RANK_REASON OpenAI published a paper detailing their methodology and findings on evaluating political bias in LLMs, including results for their latest models.
