Overview
- OpenAI’s internal study reports GPT-5 Instant and GPT-5 Thinking score about 30% lower on measured political bias than GPT-4o and o3 when handling contentious prompts.
- The Model Behavior division created a five-axis rubric—user invalidation, escalation, personal political expression, asymmetric coverage, and unwarranted refusals—scored by an automated LLM grader.
- The stress test comprised roughly 500 prompts across about 100 topics, each framed from five ideological perspectives to probe how charged wording affects responses.
- OpenAI says fewer than 0.01% of real-world ChatGPT outputs display political bias, with strongly charged liberal prompts exerting the largest pull on objectivity across model families.
- The company invited external use of its framework, but academics such as Daniel Kang call for independent audits, and critics note that key datasets and specific prompts have not been fully disclosed.