Are AI chatbots biased?

Yes, AI chatbots can exhibit various forms of bias, and this is a well-documented issue in AI research. Let me search for current information to give you a comprehensive and up-to-date answer.Yes, AI chatbots are demonstrably biased in multiple ways, according to a growing body of research. Here's a breakdown of the key findings:

Political Bias

Multiple studies have found that popular chatbots lean left politically. A 2026 Washington Post investigation found that chatbots have clear political leanings that can conflict with promises made by the companies behind them, with the model that powers ChatGPT answering nearly every question exclusively with left-leaning arguments. Specifically, OpenAI answered prompts with left-leaning responses 80% of the time while providing both sides to a question just 17% of the time, with only 3% of responses having just a right-leaning answer.

This isn't an isolated finding—the Post's results come after several previous academic studies found that AI models powering chatbots tend to favor left-leaning positions. A Danish study similarly found that all of the tested chatbots place themselves at the center or to the left of center on the political spectrum. A broader analysis of 43 language models from multiple countries found that most models leaned center-left or left, though model scale and openness were not strong predictors—pointing instead toward alignment strategy and institutional context.

Why This Happens

The bias doesn't necessarily stem from intentional programming. As one University of Michigan researcher explained, most chatbots 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, while companies write system instructions that guide their chatbots' responses. Importantly, the data that shapes AI models tends to reflect the values of Western, educated, industrialized, rich and democratic people.

Bias Affects Users—Even Unintentionally

A Yale study found something particularly concerning: while prior research showed that AI content prompted to be persuasive can shift opinions, this study provided evidence that the same is true of content not intended to change minds, such as summaries chatbots produce in response to simple factual queries. This happens because this unintended power to persuade is caused by latent biases introduced during training of the large language models that drive chatbots' core capabilities.

Bias Against Vulnerable Users

Beyond politics, MIT researchers found bias that disproportionately harms certain groups. Their study of state-of-the-art AI chatbots—including GPT-4, Claude 3 Opus, and Llama 3—found they 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. Even more troubling, these models also refuse to answer questions at higher rates for these users, and in some cases respond with condescending or patronizing language.

Bias in Product Recommendations

Bias also shows up in commercial contexts. A UC San Diego study found chatbot-generated review summaries increase users' purchase likelihood by 32% compared to original human reviews, primarily due to positive bias introduced by large language models. The researchers discovered that LLM-generated summaries changed the sentiments of the reviews they summarized in 26.5% of cases.

Other Documented Bias Types

Research also documents additional bias categories: gender bias—such as reinforcing stereotypical assumptions about occupations—and racial bias, reflected in prejudices against certain ethnic groups, with these biases largely stemming from training data and potentially amplified during both training and inference.

Important Caveats

Bias isn't necessarily fixed or permanent. As one analysis noted, evidence does not show that bias is stable—models change often, system prompts change, retrieval sources change, and political pressure changes, meaning a test from one month may not describe a model months later, which is why audits should be repeated, public, and model-specific.

Interestingly, research suggests users actually prefer balanced responses over biased ones. A Stanford study found people preferred neutral answers, even over answers that matched their own party preference, and really liked when models put in effort to describe different arguments.

Bottom line: Yes, AI chatbots exhibit measurable bias—political, socioeconomic, and otherwise—rooted in training data and design choices rather than deliberate intent. This is an active, evolving area of research, and the degree of bias varies by model, topic, and even the language used to query it.