A Washington Post investigation into six leading AI models found a clear pattern: most major AI chatbots still lean left when asked political questions. The result is especially notable because the pattern also appeared in systems presented as conservative alternatives.
The findings add weight to earlier studies showing that chatbots tend to respond with a left-leaning slant. They also suggest that efforts around "anti-woke" AI have not yet changed how these systems answer many political prompts.
The strongest left-leaning results came from OpenAI and Deepseek
OpenAI's GPT-5.5 produced the most one-sided answers in the Washington Post investigation. Eighty percent of its responses contained only left-leaning arguments, while just one answer presented an exclusively right-leaning position.
The model supported higher taxes on the wealthy and a single-payer healthcare system, among other positions. Those examples matter because they show the investigation was not only measuring tone. It was looking at whether the substance of a response lined up with arguments commonly categorized as left-leaning or right-leaning.
Deepseek's V4 Pro followed closely. It gave exclusively left-leaning answers in 70 percent of cases. Both OpenAI's GPT-5.5 and Deepseek's V4 Pro argued against the death penalty, even though a majority of Americans have supported it for decades, according to Gallup.
The key point is not that every answer from these models was identical. It is that, across the tested political questions, the investigation found a repeated tendency for the models to provide only one side of the argument, and that side was more often left-leaning.
Conservative branding did not guarantee conservative answers
The investigation also looked at models associated with conservative or anti-"woke" positioning. That part of the test is important because it separates marketing from observed behavior.
xAI's Grok 4.3, promoted by Elon Musk as a "truth-seeking" and anti-"woke" chatbot, produced more right-leaning answers than any other model tested. Even so, it still gave exclusively left-leaning responses more often.
The source article points to one likely explanation: Grok may have been trained on the same data as other chatbots, or even on their outputs. That would help explain why a model presented as different might still reproduce a similar political pattern across many topics.
Grok's behavior also shows that model alignment can be steered. In the Washington Post test, it took an exclusively right-leaning position on trans rights, a stance that the source says matches Musk's public position. That suggests intervention in the model's output at least on certain topics.
Gab's Arya showed a related mismatch between positioning and results. Gab describes Arya as "built with Christian values and conservative principles." Yet in the Washington Post investigation, Arya responded with a left-leaning argument twelve times more often than a right-leaning one.
Gemini 3.1 Pro was the main exception
Google's Gemini 3.1 Pro stood out from the other AI models tested. It presented both sides 93 percent of the time. Only seven percent of its answers were exclusively left-leaning, and it never gave an exclusively right-leaning response.
That does not mean Gemini 3.1 Pro avoided controversial answers. When asked whether the U.S. should use its military to conquer new territory, it was the only model that offered an argument in favor of expansion, saying it could strengthen the U.S. economy.
Still, the broader result was that Gemini 3.1 Pro most often framed political questions by presenting competing perspectives. In a test focused on whether AI chatbots lean left or right, that made it the balanced outlier.
Why the left-right label has limits
The Washington Post made its full code and supplementary analysis available on GitHub, according to the source article. That transparency matters because political classification is difficult, especially when answers are reduced to categories such as left, right, or both sides.
The source also raises a caution about the method. Sorting AI answers into "left" and "right" may be too simple for some topics. On certain questions in the test, right-leaning positions conflict with scientific consensus or universal human rights.
That creates a difficult design problem for chatbot makers. A model can be criticized for political imbalance if it refuses to present a conservative argument. But if the requested argument would relativize facts or basic rights, presenting it as an equal perspective may create a different problem.
The investigation therefore points to two issues at once. First, many major AI chatbots still show a left-leaning pattern on political questions. Second, measuring political balance in AI is not always as simple as asking whether both sides appeared in the answer.
For users, the practical takeaway is clear: political answers from AI chatbots should be read as model outputs, not neutral ground truth. Even when a chatbot is marketed as conservative, balanced, or truth-seeking, its responses may still reflect training data, alignment choices, and topic-specific interventions.