Defining and evaluating political bias in LLMs

OpenAI introduced a new framework to test political bias in ChatGPT using about 500 prompts across 100 topics.

The company says ChatGPT is near-objective on neutral prompts, shows moderate bias on harder politically charged prompts, and GPT-5 performs better than earlier models.

OpenAI estimates political bias appears in less than 0.01% of production responses, suggesting the issue is rare in real traffic.

The work matters because political neutrality affects user trust and how reliable AI is as a learning tool.