Are AI chatbots biased?

Are AI chatbots biased? I'll search for recent research and evidence on chatbot bias to answer.

Yes, AI chatbots can be biased — but "biased" needs context.

In research, bias means a systematic tendency in outputs rather than a conscious political intention. It does not mean every chatbot always gives the same slanted answer. It means that across many prompts, a recurring pattern appears.

1. What the research shows

Political bias is the most studied:

•The best recent evidence points to a left-of-center tilt in many major models on U.S. and European political questions.

•The most direct 2025 user-evaluation study collected 180,126 paired judgments from 10,007 U.S. respondents, testing 24 large language models on 30 political topics. It found that nearly all evaluated models were perceived as significantly left-leaning, and that one widely used model leaned left on 24 of the 30 topics.

•A March 2026 benchmark of eight prominent models found that seven leaned left and one, Grok, leaned right.

•A June 2026 Washington Post test found ChatGPT's model gave almost exclusively left-leaning arguments while Google Gemini usually presented both sides in more than 90 percent of its answers.

•In Denmark, researchers from the University of Copenhagen found "Overall, all of the tested chatbots place themselves at the center or to the left of center on the political spectrum". When asked to recommend parties to fictitious voters, recommendations were far from evenly distributed - some parties were recommended disproportionately often while others were not suggested at all as first choice by some models.

Not all studies find strong bias for every model. A Chalmers University of Technology review of nearly 150,000 AI-generated texts showed that the chatbots' responses did not benefit any single parliamentary party, but usually placed themselves in the middle of Swedish politics. Only one language model stood out significantly: Grok.

Other types of bias are also well-documented:

•Gender bias where, for example, doctors are assumed male and nurses female

•Racial bias, which might cause identical text to receive different treatments based on the perception of the author's race

•Cultural bias, which might favour Western perspectives over others

•Health and dialect bias: Models were found to stereotype users with health conditions and to show increases in stereotyping for non-American/British dialects.

Historical examples include:

•Microsoft's chatbot Tay, which within 24 hours started tweeting offensive and racist content

•Bloomberg's 2023 analysis of 5,000 images from Stable Diffusion found that the model amplified racial and gender stereotypes well beyond their representation in the actual workforce

Crucially, bias can influence users even when not intended. A Yale-led PNAS Nexus experiment with 1,912 participants found default GPT-4o historical summaries shifted opinion in a liberal direction compared with Wikipedia summaries, and research warns there is no truly neutral AI in the sense that neutrality itself is hard to define in politics.

2. Why chatbots are biased

Large Language Models like ChatGPT, Claude, Gemini, and DeepSeek are all biased. In fact, they can be said to be born biased.

Reasons include:

•Training data reflects human society. AI chatbots like ChatGPT are trained on massive troves of documents created by human beings, some of whom are antisemitic, racist, and/or sexist. All human biases are within the generative database.

•These systems inherit and amplify the biases embedded in human-generated training data.

•Bias is fundamentally unavoidable because they learn from human-written text which contains centuries of accumulated societal biases.

•Pattern learning: Bias in AI models arises from overlapping mechanisms that amplify one another. Training data patterns shape predictions: if the word she appears with nurse ten times more often than with engineer, the model is likely to complete the phrase "The senior engineer said that ___" with he.

•The "averaging" effect: When an AI model creates an averaged response based on the available documents in the training data, it is basically stereotyping.

•Alignment and product policy: Bias can enter at training data, filtering, post-training alignment, reward modeling, system prompts, and user wording — not just raw data.

Researchers distinguish two levels: Intrinsic bias becomes encoded in the model's internal representations during training, while extrinsic bias emerges when models perform differently across demographic groups in real-world applications.

3. How bias manifests

•Not usually as blatant factual errors. Bias may appear less as a factual error than as a choice of emphasis, vocabulary, cause, blame, moral lesson, or omitted counterargument.

•The harm is of two main types: Representational harm when models misrepresent groups through negative portrayals, and allocational harm when models influence resource distribution like employment or healthcare.

4. Is it fixable?

Studies show bias is not a single fixed trait. It varies by model, prompt, topic, language, training method, and whether the model is forced to answer.

•A neutrality prompt reduced perceived slant in the large 2025 user study, suggesting some tilt can be reduced by instruction design, although not necessarily eliminated.

•Different products today behave very differently — as the Washington Post test showed, some models present mostly one-sided arguments while others present both sides most of the time.

The UK Government's AI Insights report notes that because bias management must be continuous, effective mitigation requires embedding bias checks at every stage from development through production monitoring, not as a one-time audit.

> Bottom line: Yes — current AI chatbots do show measurable biases (political, gender, racial, cultural and others) that reflect and sometimes amplify human biases in their training data. The strongest and most replicated finding is a tendency for many popular English-language models to lean center-left on contested political and social issues in the US and Europe, but the size and direction of bias depends heavily on the specific model, prompt, and topic. Critical use — asking for balanced perspectives, comparing sources, and not treating a chatbot as a neutral oracle — remains essential.

Citations (4)

1etcjournal.comNews and media
AI Chatbots Are Liberal-Left ‘Biased’
Backs: “I'll search for recent research and evidence on chatbot bias to answer. Yes, AI chatbots can be biased — but "biased" needs context. In research, bias means a systematic tendency in outputs rather than a conscious political intention.”
2phys.orgNews and mediaquality 0.85
Chatbots show political bias and steer voters toward some parties, analysis finds
Backs: “A June 2026 Washington Post test found ChatGPT's model gave almost exclusively left-leaning arguments while Google Gemini usually presented both sides in more than 90 percent of its answers. In Denmark, researchers from the University of Copenhagen found "Overall, all of the tested chatbots place themselves at the…”
3gov.ukGovernment
AI Insights: Large language models (LLMs) Bias (HTML)
Backs: “Only one language model stood out significantly: Grok. Other types of bias are also well-documented:”
4counterpunch.orgNews and mediaquality 0.46
Is AI Born Biased?
Backs: “Why chatbots are biased. Large Language Models like ChatGPT, Claude, Gemini, and DeepSeek are all biased.”