80% of OpenAI’s GPT-5.5 Answers Leaned Left in Political Bias Test

For AI developers and researchers, the political leanings of large language models are more than a PR concern—they reflect the underlying data, alignment processes, and corporate governance shaping these systems. The Washington Post's test of six popular chatbots reveals a pervasive leftward tilt that could undermine the credibility of AI as an objective knowledge tool. Understanding why GPT-5.5 leans left 80% of the time and how Gemini stays balanced is critical to building trustworthy AI.

A new investigation by The Washington Post has uncovered a pronounced left-wing bias in the responses of popular AI chatbots, raising serious questions about the political neutrality of large language models that are increasingly used as sources of information and decision support. The Post tested six leading chatbots—including models from OpenAI, Google, Anthropic, DeepSeek, Gab, and Elon Musk’s xAI—by posing researcher-designed questions on 30 contentious political topics such as affirmative action, universal basic income, mass deportation, and tariffs. A human reporter then scored the answers for how much they leaned left or right. The results were stark: OpenAI’s GPT-5.5 answered a full 80% of the queries with a leftist slant, far exceeding any other model in one-sidedness. Google’s Gemini proved to be the most balanced, presenting both sides of the argument in over 90% of its responses. Claude Opus 4.8 from Anthropic achieved balance 57% of the time, while the remaining 43% of answers leaned left and none leaned right. Elon Musk’s Grok 4.3, despite its positioning as a counter to ‘politically correct’ AI, gave right-leaning responses only about a third of the time—the highest such share in the test, but still a minority. Even Gab’s chatbot, associated with a conservative social network, produced left-leaning answers half of the time, underscoring the pervasive tilt. The Chinese AI company DeepSeek’s model also mostly leaned left.

“

Claude Opus 4.8 from Anthropic achieved balance 57% of the time, while the remaining 43% of answers leaned left and none leaned right.

The implications of this bias extend far beyond partisan debate. AI chatbots are rapidly becoming integral to search engines, enterprise knowledge bases, and even legal and medical advisory tools. If these models systematically favor one side of the political spectrum, they risk distorting public understanding, undermining trust in AI systems, and potentially violating emerging regulatory standards. The fact that Gemini could achieve such high neutrality suggests that bias is not an inevitable consequence of language model architecture but rather reflects deliberate design choices in data curation, reinforcement learning from human feedback (RLHF), and corporate content policies. For AI researchers and developers, the test serves as a reminder that value alignment is as critical as technical performance.

The sources of the bias are multifaceted. Training data drawn disproportionately from left-leaning internet forums, academic publications, and mainstream media can embed an ideological skew. The RLHF process, where human annotators score model outputs, may introduce their own political perspectives, especially if the annotator pool lacks ideological diversity. Additionally, companies’ commitments to safety and inclusivity sometimes translate into avoiding language that could be perceived as offensive or exclusionary, inadvertently steering models toward progressive framings. The WaPo test exposes the outcome: models that, by default, advocate for specific policy positions—like opposing tariffs or supporting pathways to citizenship for undocumented immigrants—rather than neutrally exploring trade-offs.