Major AI Chatbots Show Left-Leaning Bias on Political Questions

A Washington Post investigation has revealed that many major AI chatbots tend to provide left-leaning responses to political questions. Notably, a wide range of widely-used chatbots, including models claiming to be "anti-woke," have not escaped this trend, which has emerged as a key finding of the investigation.

Discussions surrounding the political bias of AI chatbots have been ongoing for some time. As generative AI becomes more widespread, increasing numbers of people are using it as an alternative to search engines and information gathering tools, and interest in the political nuances contained in its responses is also growing. While developers often emphasize the neutrality of their models, there exists a reality where it is difficult for external parties to verify whether actual outputs lean in a particular direction.

Looking at the survey results in detail, OpenAI's GPT-5.5 returned only left-leaning arguments in response to political questions in as much as 80% of cases. Meanwhile, Grok, which Elon Musk positioned as "anti-woke," also reportedly showed a frequency of left-leaning responses that exceeded right-leaning ones. These results demonstrate a discrepancy between the positioning espoused by developers and their actual outputs.

A markedly different result came from Google's Gemini 3.1 Pro. This model presented arguments from both left and right perspectives in 93% of cases in response to political questions, displaying notably greater balance compared to other models. However, it should be noted that whether this "presentation of both perspectives" constitutes substantive neutrality may vary depending on the design of questions and evaluation criteria.

An important implication shown by this survey is that the brand image of an AI model does not necessarily align with its actual output tendencies. Labels such as "neutrality" or "anti-woke" may be used for marketing appeal, but whether actual responses possess the same characteristics is a separate matter. As users increasingly employ AI as an information source, the impact of such biases is far from negligible.

AI behavior on political and social topics is deeply connected to the design of the model's training data and mechanisms for adjusting responses, such as RLHF (reinforcement learning from human feedback). The specific details of what data and guidelines developers use to adjust their models are often not disclosed. For this reason, continuous verification by external organizations is positioned as playing a crucial role in ensuring AI transparency.

Points to watch going forward include whether each company will adjust its model calibration policy in light of these survey results. There also remains the question of how to establish evaluation standards for political neutrality itself. As AI takes on an increasingly prominent role as social infrastructure, the methods for verifying and disclosing bias are expected to continue being questioned, intertwined with regulatory discussions.