GPT-5.5 and Claude 4.8 far more balanced than WaPo bias study claims

A recent Washington Post study went viral, claiming ChatGPT exhibits massive left-wing bias 80% of the time. Researcher John-Clark Levin replicated the experiment for GPT-5.5 and Claude Opus 4.8 and found three glaring problems. First, the study forced AIs to answer political questions in ≤30 words at a 9th-grade reading level—conditions no real user employs. Second, several 'right-wing' positions (e.g., supporting military conquest for resources, banning labor unions) are not actually mainstream conservative views. Third, factual positions like tariff skepticism were arbitrarily scored as left-leaning, even though pre-Trump Republicans held that view. Under those artificial constraints, Levin got similar results (80% left-only for GPT-5.5).

But when Levin tested under realistic conditions—removing the word limit and simplifying language—GPT-5.5's left-only responses fell to 62%, then to 34% with the full system prompt removed. After excluding fringe questions that don't reflect genuine political debates (requiring ≥30% support from both parties), the left-only share plummeted to 15.8%. Claude Opus 4.8 fared even better: from WaPo's 43.3% left-only, realistic conditions dropped it to 9.3%, and excluding fringe questions produced 0%—Claude considered both left and right arguments 100% of the time. The study's claimed effect essentially disappeared, though a slight left-leaning tendency remains. The finding undermines President Trump's narrative that AI labs are pushing radical leftism, reducing political ammunition that could force models to skew outputs.