Are AI chatbots biased?

Yes — research consistently shows that AI chatbots exhibit bias, and this shows up in several distinct ways:

1. Political bias

Multiple independent studies and media tests have found that major chatbots lean in a particular political direction. A Washington Post test found that OpenAI's ChatGPT showed left-leaning bias 80% of the time when answering political questions, and even Grok, widely used by conservatives on X, still had a slight liberal bias, with 40% of its responses left-leaning, 33% right-leaning and 27% balanced. Google's Gemini behaved differently — Gemini mostly took a both-sides approach, offering both left and right positions in more than 90 percent of its answers.

This pattern isn't limited to one test. A large-scale 2025 user study found that a direct 2025 user-evaluation study collected 180,126 paired judgments from 10,007 U.S. respondents across 24 large language models on 30 political topics, finding that nearly all evaluated models were perceived as significantly left-leaning. The bias even seems to creep into the "alignment" machinery itself — one study found that reward models showed left-leaning bias even when trained on factual data, with larger models showing more bias.

Researchers note the bias originates largely from training choices: most models are trained on large collections of text scraped from the internet, but companies can choose what data to include, and AI firms also hire workers to refine what their models say by scoring which responses are considered better.

2. Bias can actually sway people's opinions

This isn't just academic — biased chatbots can change how people think in real time. A University of Washington study found that Democrats and Republicans were both more likely to lean in the direction of the biased chatbot they talked with than those who interacted with the base model, with people from both parties leaning further left after talking with a liberal-biased system. One researcher noted: "These models are biased from the get-go, and it's super easy to make them more biased."

More strikingly, large-scale election research published in Nature and Science found chatbots can be even more persuasive than traditional political ads. A politically biased AI model was more effective than political advertisements at nudging both Democrats and Republicans toward the opposing party's candidates — swaying opinions by citing facts and evidence, though the most persuasive models said the most untrue things. In the U.S. test, the pro-Harris AI model moved likely Trump voters 3.9 points toward Harris on a 100-point scale — an effect roughly four times larger than traditional ads — while the pro-Trump model moved likely Harris voters 1.51 points toward Trump. In other countries the effect was even larger: in Canada and Poland, chatbots moved opposition voters' attitudes and voting intentions by about 10 percentage points.

3. Bias beyond politics — gender, race, and socioeconomic status

Bias extends well past political topics. AI chatbots may exhibit gender bias, such as reinforcing stereotypical assumptions about men and women's occupations, as well as racial bias reflected in prejudices against certain ethnic groups.

A 2025 Stanford-reported study found troubling patterns in hiring contexts: when researchers asked AI to generate resumes for hypothetical female candidates, it consistently portrayed them as younger and less experienced than male counterparts, and when asked to rate those resumes, it gave higher scores to older men than older women even with identical qualifications.

Healthcare is another area of concern. A Cedars-Sinai study found that leading large language models, including Claude, ChatGPT, and Gemini, tend to generate less effective treatment recommendations when a patient's race, whether implied or explicitly stated, is African American.

There's also evidence of bias based on how a user presents themselves. An MIT study found that state-of-the-art AI chatbots — including GPT-4, Claude 3 Opus, and Llama 3 — sometimes provide less-accurate and less-truthful responses to users who have lower English proficiency, less formal education, or who originate from outside the United States. In one striking example, Claude 3 Opus refused to answer nearly 11% of questions for less-educated, non-native English-speaking users compared to just 3.6% for the control group, and responded with condescending or mocking language 43.7% of the time for less-educated users versus less than 1% for highly educated users.

Why this happens

At a root level, bias stems from training data and human feedback: AI models' biases largely stem from the data on which they are trained, and can be amplified during both training and inference.

The good news: researchers have found bias can be reduced. One Stanford-affiliated study showed that with just a small tweak, many models can be prompted to take a more neutral stance that more users trust — meaning developers (and even users, through prompting) have real tools to mitigate these effects, even if they can't eliminate them entirely.

Bottom line: Yes, AI chatbots are biased — in political leaning (most tilt left, though to varying degrees), in demographic stereotypes (gender, race, age), and in how they treat users based on perceived education or language background. These biases are measurable, can meaningfully influence real-world opinions, and are an active area of research and mitigation effort.