OpenAI published a human-level evaluation of political bias, built on 500 prompts across 100 subjects and five measurable factors: user invalidation, escalation, personal political voice, asymmetric coverage, and unjustified refusals.
Models are nearly objective on even-handed or weakly skewed prompts and exhibit moderate bias on strongly emotive prompts; GPT-5 instant thinking and GPT-5 thinking reduce measurable bias by approximately 30% compared to previous models.
To assess impact in the outside world, OpenAI applied the same method to production traffic and estimated that less than 0.01% of responses exhibit any political bias.
Bias, when it does occur, most frequently appears as the model expressing opinions under its own guise, one-sidedness, or amplification of a user’s inflammatory language.
OpenAI is looking forward to further development, hoping to be stronger while gauging progress with automatic and interpretable tests.
You may also want to check out some of our other recent updates.
Wanna know what’s trending online every day? Subscribe to Vavoza Insider to access the latest business and marketing insights, news, and trends daily with unmatched speed and conciseness! 🗞️
