AI chatbots are at the center of a controversy that blends technology and politics in ways few people expected.
When President Donald Trump signed an executive order mandating that AIs be neutral and nonpartisan tools, the debate over political bias in these technologies moved from the realm of speculation to a concrete, urgent issue with real implications for millions of people. The move also triggered reactions from Democrats, who began to worry that AI could start leaning right under political pressure.
But do chatbots actually have political bias?
The Washington Post decided to go beyond the controversy and put that question to a real test, evaluating the leading AI models available today with political questions calibrated by researchers from universities like Dartmouth and Stanford. The models behind ChatGPT from OpenAI, Gemini from Google, as well as Anthropic, DeepSeek, xAI, and Gab were all assessed.
The results were, to say the least, surprising 👀
Among the most striking findings, the model behind ChatGPT answered nearly all questions with only left-leaning arguments, presenting a right-leaning position just once, totaling around 80% of responses with a progressive slant. Meanwhile, Google’s Gemini was the only model to present both sides in more than 90% of its responses.
And there’s more: even Grok, from Elon Musk’s xAI, which is marketed as an anti-woke alternative focused on the pursuit of truth, also cited more progressive arguments on average overall.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
In this article, you’ll learn how the test was conducted, what the numbers reveal about each model, what the companies said about the results, and why total neutrality in AI might, in practice, be an illusion that’s incredibly hard to achieve 🤖
How the test was conducted
The methodology behind the Washington Post investigation was anything but improvised. The newspaper based its approach on research published last year by the lab of Sean Westwood, director of the Polarization Research Lab at Dartmouth College, in partnership with researchers from Stanford University. That study developed more than two dozen political questions designed to reflect the kind of things everyday people might ask a chatbot.
Each model was asked to answer the questions in no more than 30 words, with no personalization settings enabled. A reporter reviewed the responses to classify whether they contained a left-leaning position, a right-leaning position, or both. Political topics rarely split cleanly between parties, but the questions covered a wide range of subjects, and the newspaper checked whether the models were consistent in their responses by repeating each question five times to ensure data reliability.
This combined quantitative and qualitative analysis is what gives the results scientific weight and makes them hard to dismiss, regardless of the reader’s political leanings. According to Westwood, understanding the positions these tools amplify matters because they are becoming increasingly influential as more people use them to make sense of the world and news events. As he put it, these AI tools are not presenting, on average, a truly neutral representation of nuanced political debates.
It’s worth noting that this type of study isn’t entirely new in the field of bias research in artificial intelligence. Several earlier academic studies had already found that AI models tend to favor left-leaning positions. But the timing of this particular investigation carried extra weight because of the political context in the United States. With Trump’s executive order putting the topic on the public agenda and tech companies being pressured to take a stance, the data gathered by the Washington Post arrived at a moment when the discussion about political tendencies in AI systems is no longer a debate confined to academics or enthusiasts. It’s a conversation that has hit the mainstream.
What the numbers reveal about each model
The result that drew the most attention was from the model behind ChatGPT, developed by OpenAI, which gave the most slanted responses of all. Around 80% of them presented only left-leaning arguments. The model argued, for example, in favor of abolishing the Electoral College in favor of choosing the president by popular vote, raising taxes on the wealthy, and adopting a single-payer healthcare system. This doesn’t necessarily mean the model was programmed to defend a specific ideology, but it does indicate that the training data, combined with OpenAI’s fine-tuning choices, resulted in a response pattern that leans in one direction. This is exactly the kind of bias that concerns researchers, because it operates subtly, often invisible to the average user who receives an answer and accepts it as neutral.
The AI from Chinese company DeepSeek came in right behind and also leaned left in its responses. Both it and the OpenAI model argued against the death penalty, which, according to Gallup, has been consistently supported by a majority of Americans for decades. This contrast between the models’ positions and majority public opinion illustrates well how bias can slip through even on topics where there is a relatively clear social consensus.
Google’s Gemini performed quite differently in this comparison. It was the only model to present both sides of the debate in more than 90% of responses, which makes it stand out when the criterion is political balance. It even offered arguments from both sides on a question about whether the United States should use its military power to conquer new territories in search of resources. No other model presented an argument in favor of conquest. This doesn’t mean Gemini is perfect or completely free of bias, but it suggests that Google adopted different alignment strategies, investing more in techniques that push the model to present multiple perspectives before wrapping up a response on sensitive topics.
And then comes what might be the most curious data point of all: Grok, from Elon Musk’s xAI, frequently positioned in the market as an anti-woke alternative to other chatbots, gave more right-leaning responses than any other model in the test, but most of the time, it still offered an entirely left-leaning position. This result is especially significant because it directly contradicts the product’s marketing. Musk has been quite vocal about what he calls left-wing bias in competing systems, and Grok was created, in part, as an answer to that perception. Seeing the model exhibit a pattern similar to what it criticizes is, at the very least, a data point that deserves some reflection on just how difficult it is, in practice, to build a truly neutral system on politically charged topics 🤔
Interestingly, the right-leaning social network Gab offers a model called Arya, described as built with Christian values and conservative principles. Yet, in the Washington Post tests, it responded with a left-leaning argument 12 times more often than with a right-leaning one.
What the companies said
Faced with the results, the companies involved were asked for comment, and the responses followed a familiar pattern in these situations. Google spokesperson Lauren Fine stated that Gemini was designed to provide balanced responses that don’t favor any political ideology, and that the company was unable to reproduce the one-sided responses that appeared in the newspaper’s tests.
Anthropic spokesperson Michael Aciman said the company trains Claude to treat different political viewpoints equally and that it extensively tests for bias before each model launch. According to him, the Post’s tests don’t reflect how most people use the company’s products, since Claude generally has more room to include context when discussing politics.
OpenAI, through spokesperson Liz Bourgeois, stated that ChatGPT was built to be objective by default and to help people explore ideas from different perspectives, and that the company works to measure and reduce political bias. OpenAI also said it was unable to replicate the study’s findings. SpaceX, DeepSeek, and Gab did not respond to requests for comment. It’s worth remembering that OpenAI CEO Sam Altman himself said back in 2023 that the company would try to make the default version of ChatGPT as neutral as possible, but that the solution lies in customization, because the word neutral means different things to different people.
Why total neutrality in AI is nearly impossible
Understanding why chatbots have political bias requires understanding how these systems are built. Large language models, known as LLMs, are trained on massive volumes of text collected from the internet, books, articles, forums, and other sources. The companies choose which data to include, and that dataset reflects the world as it is, with all its contradictions and cultural leanings. When a model learns from that material, it inevitably absorbs the patterns present in it, including political patterns. There is no politically neutral training corpus because human content itself is not neutral.
Beyond the initial training, models go through a fine-tuning process with human feedback, where evaluators indicate which responses are better or worse. Companies also write system instructions that guide chatbot behavior. This process, however well-intentioned, introduces yet another layer of subjectivity, because the evaluators also have their own opinions and leanings. Ceren Budak, a professor at the University of Michigan who studies how technologies interact with political polarization, notes that the data shaping AI models tends to reflect the values of people who are Western, educated, industrialized, rich, and democratic. In her view, it would be helpful to have more clarity about the companies’ current value systems, so that users know what they’re working with.
Budak also highlights an important point: with AI, tech companies are taking on a more active role, because the products generate political speech directly, rather than simply curating human speech the way social media platforms do. Even people who never talk to chatbots about politics end up exposed to AI-generated text in online content and through other channels.
The most recent trends in AI alignment attempt to address this problem through techniques that force the model to present multiple perspectives on controversial topics. The Gemini result in this study suggests that these approaches can work to a certain extent, but there’s still a long way to go. Andrew Hall, a Stanford researcher who participated in the original study, said he was surprised that not all major chatbots responded as neutrally as Gemini, expecting that the other models would have already reached that benchmark. Many scholars argue that political neutrality is actually impossible, since even middle-ground positions are positions in themselves and tend to benefit the stronger side. By that logic, Google’s both-sides approach is, in itself, a political choice. The fundamental issue is that defining what’s neutral is already, in itself, a political act 🤖
The impact of this discussion on the future of AI
The debate over political bias in chatbots isn’t going away. In fact, everything points to it intensifying in the coming months and years as these systems become increasingly present in everyday life. Few Americans use AI directly to understand politics, but according to a March survey from the Polarization Research Lab, nearly half use AI occasionally to keep up with the news. When millions of people turn to a chatbot for information, that system’s bias has an influence potential that goes beyond any traditional media outlet, simply because the scale is unmatched and the interaction is personalized, creating a sense of trust that’s hard to question.
Hall, from Stanford, also points out that AI companies need to deal with different categories of questions. Some are objective and factual, like what is the speed of light, which chatbots answer with ease. But most political questions don’t have that characteristic, since there is no single truth. You have to take the facts and layer values on top of them, and that’s where the difficulty lives. Westwood, from Dartmouth, observes that both Democrats and Republicans don’t trust AI to be neutral and keep it away from their voting decisions, which is one of the few points where both sides agree in today’s political landscape.
Trump’s executive order, regardless of what anyone thinks about it, put a finger on a real wound and brought to the surface a conversation the tech industry needed to have more openly. The companies developing these systems carry a significant responsibility regarding how their models behave in political contexts, and studies like the one conducted by the Washington Post with researchers from Dartmouth and Stanford are important tools for keeping that responsibility visible and auditable.
What becomes clear after analyzing all this data is that the discussion about political tendencies in artificial intelligence systems is not a simple issue to solve with a software update or a well-crafted corporate statement. It demands a continuous, multidisciplinary, and honest effort from those who develop these technologies, and a critical, informed stance from those who use them. Chatbots are extraordinarily powerful tools, and understanding their political limitations is just as important as leveraging their capabilities. After all, an AI that appears neutral but isn’t can be more dangerous than one that openly acknowledges its limitations 💡
