OpenAI finds 30% drop in ChatGPT political bias

OpenAI released new findings on how it measures and addresses political bias in ChatGPT.

The company tested its models with about 500 prompts across 100 topics, using a framework to detect five types of political bias, including personal opinions and one-sided coverage.

GPT-5 instant and GPT-5 thinking showed a 30% reduction in bias compared to previous versions when responding to emotionally charged prompts.

OpenAI estimated that less than 0.01% of ChatGPT’s real-world outputs showed any political bias, based on its internal evaluation of production traffic.

The company said it will keep working to improve model objectivity, especially for sensitive topics, and noted that political and ideological bias in language models remains an open research issue.

🔗 Source: OpenAI

🧠 Food for thought

Implications, context, and why it matters.

OpenAI bias claims lack a reproducibility standard in AI safety

OpenAI says bias fell 30% with under 0.01% biased replies 1. It has shared no public prompt set or full rubric or independent audit that lets others test if the 30% gain reflects neutrality or masked signals via prompt engineering (the practice of wording instructions to steer model outputs).
Stanford researchers ran a 30-question test across 24 large language models (LLMs) from eight firms, with over 10,000 people rating answers 2. Earlier work found both Republicans and Democrats perceived a left tilt, with OpenAI models showing four times the slant of Google models 2.

EU AI Act rules spur demand for outside bias audits for GPAI and high-risk providers

Article 10 of the European Union’s AI Act (a comprehensive regulation governing AI) requires high-risk AI to detect and reduce bias 3 4. Teams must assess impacts on fundamental rights and, when strictly necessary, log special-category data with safeguards (special-category data includes traits like health status or political opinions).
Systemic risk GPAI models exceed the ≥10^25 floating point operations (FLOPs) training threshold, a proxy for large compute 4. They must run systemic risk checks and incident reporting, with fines up to €35 million or 7% of global turnover for the most serious breaches 4.
Independent auditors can provide continuous monitoring with AI and human ratings 3 5 4. This supports Article 10 data governance plus GPAI documentation, risk duties under Articles 53–55. Codes of Practice are due in April 2025 to guide implementation 5.