Washington Post Study: ChatGPT and Claude Lean Left More Than Grok and Gemini

Six of the most popular AI chatbots answer political questions with a clear left-leaning tilt, according to an analysis conducted by Washington Post journalists in collaboration with researchers from Dartmouth College and Stanford University. The exception is Google's Gemini, which presents both sides of the argument in the vast majority of cases, and xAI's Grok, the only model that formulates right-leaning arguments more often than the other tested models.

The authors of the analysis asked the models a series of questions on contentious political topics, including taxes, healthcare, the death penalty and transgender rights, then classified the answers according to whether they presented only one side of the argument or balanced arguments from both political options. The questions and methodology were developed by researchers from the Polarization Research Lab at Dartmouth College and a team from Stanford University.

Who fared worst

The most one-sided model turned out to be the OpenAI system powering ChatGPT, which presented exclusively left-leaning arguments in 80 percent of answers, and presented both sides of the argument in only slightly over ten percent of cases. China's Deepseek V4 Pro performed similarly, with 70 percent of its answers one-sidedly left-leaning. Anthropic's Claude fared somewhat better than ChatGPT, but still limited itself to one side's arguments in 43 percent of answers.

Gemini and Grok at opposite ends

Google's Gemini 3.1 Pro turned out to be the most balanced of the tested models, presenting arguments from both sides of the political divide in 93 percent of answers. Only 7 percent of Gemini's answers were classified as exclusively left-leaning. A Google spokesperson emphasized that the model was designed not to promote any political agenda.

xAI's Grok, promoted by Elon Musk as an alternative to allegedly biased competitors, also showed a lean toward left-leaning answers over right-leaning ones, 40 percent versus 33 percent. However, of all six tested models, Grok formulated right-leaning arguments most often, making it the most balanced in terms of representing both political perspectives after Gemini.

Methodology and limitations

Researchers at the Polarization Research Lab at Dartmouth College, led by Sean Westwood, prepared a set of questions covering sensitive social and economic topics, and the models' answers were classified according to whether they contained only one perspective or balanced both positions. The Washington Post published the code and the full analysis, allowing other researchers to independently verify the methodology.

Even models marketed as less censored or more neutral still tend toward left-leaning interpretive frames on most contentious topics in practice - Sean Westwood, director of the Polarization Research Lab, Dartmouth College

The companies behind the tested models consistently deny deliberately favoring one ideological option. Representatives of OpenAI and Anthropic point out that the bias may stem from the nature of the training data rather than deliberate design decisions, and say they are continuing to work on reducing one-sidedness in responses.

Why it matters

Millions of users worldwide, including a growing number in Poland, treat AI chatbots as their first source of information for social and political questions, replacing traditional web searches. If responses systematically favor one side of the debate, this could shape public opinion on a scale previously unavailable to any single media outlet.

The issue takes on added significance in the context of the EU's AI Act and discussions about transparency for high-risk artificial intelligence systems. In Poland, the topic of algorithmic bias has already come up in warnings from the Obserwatorium Demokracji Cyfrowej (Digital Democracy Observatory) about the possible influence of algorithms on election outcomes, though analyses so far have focused mainly on social media rather than conversational chatbots.

For companies deploying AI chatbots in customer service or internal communication, the study's findings mean additional reputational risk, especially when assistants answer employee or customer questions on socially sensitive topics. AI ethics experts recommend testing one's own deployments for similar bias before production launch, rather than assuming the model provider's neutrality.