OpenAI Reports 30% Reduction in Political Bias in GPT-5 Models

OpenAI announced in a research post that its latest AI models, GPT‑5 Instant and GPT‑5 Thinking, demonstrate a 30% reduction in political bias compared to earlier versions such as GPT‑4o. The company said its evaluation framework was designed to measure objectivity in realistic user interactions across politically charged and neutral prompts.

The evaluation involved around 500 questions covering 100 topics, each written from five ideological perspectives ranging from liberal to conservative. Responses were analyzed across five measurable axes of bias—user invalidation, user escalation, personal political expression, asymmetric coverage, and political refusals. According to OpenAI, less than 0.01% of all ChatGPT responses show any signs of political bias in real-world use.

The company reported that GPT‑5 models remain largely neutral in typical conversations but can still display moderate bias when responding to emotionally charged prompts. These results were consistent across both GPT‑5 variants and indicate improved robustness under challenging conditions.

OpenAI stated that it will continue refining its models’ objectivity and invited external researchers to use its framework for independent evaluation. The improvements follow the earlier introduction of GPT‑5, which OpenAI describes as its most advanced generation of language models to date.