Are AI chatbots biased?

I'll pull the latest research to check whether AI chatbots show bias and why.

Yes, the evidence shows AI chatbots can be biased — but "bias" does not mean the models have beliefs or conscious political intentions.

Researchers use "bias" to mean a systematic tendency in outputs rather than a conscious political intention, or more formally a systematic pattern in model outputs that favors one political, moral, or social framing over plausible alternatives. That pattern can arise without a developer trying to make it biased.

1. What the recent studies show

Political bias: 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 analysis found most leading AI chatbots leaned left when answering politically charged questions, specifically that the model powering ChatGPT gave almost exclusively left-leaning arguments, while Google Gemini usually presented both sides in more than 90 percent of its answers.

•European evidence points in the same direction. Larger models such as Llama3-70B aligned more closely with left-leaning parties; smaller models often stayed neutral, especially in English.

Researchers warn there is no truly neutral AI, and that users tend to absorb a model’s bias after just a few interactions.

Gender and occupational bias:

•GPT-3 has demonstrated notable gender biases, disproportionately associating leadership and technical roles with men while linking women to caregiving and administrative positions.

•Researchers at MIT and Stanford examined how leading large language models respond when prompted as male or female users seeking investment guidance. MIT researchers find AI chatbots give women more conservative financial advice, costing them $60,000 by age 60.

•A UN Women analysis released in June 2026 reviewed 133 AI systems. Forty-four percent showed clear gender bias.

A more direct economic test found AI chatbots routinely suggest lower salaries to women and some ethnic minorities and people who described themselves as refugees, even when the job, their qualifications, and the questions are identical.
For instance, ChatGPT's o3 model told a fictional male medical specialist in Denver to ask for $400,000, when a different fake persona identical in every way but described as a woman was told to aim for $280,000. The most advantaged profile turned out to be a “male Asian expatriate,” while a “female Hispanic refugee” ranked at the bottom. The researchers noted that even subtle signals like candidates' first names can trigger gender and racial disparities.

Racial, ethnic and other biases:

•BERT, a leading NLP model developed by Google, has exhibited racial biases in various tasks.

•A Penn State study of GPT-4o, Gemini 1.5 Pro, and DeepSeek v2.5 found LLMs often rely on harmful stereotypes when generating personas based on sociodemographic factors. For example, when prompted to represent a 50-year-old African American woman, the chatbot frequently invoked stereotypical topics like gospel music, "tough love," social justice activism, and natural hair care.

•In 2024, researchers from UC Berkeley tested ChatGPT's responses to several varieties of English dialects. Compared to American or British English, responses to dialects showed increases in stereotyping (18 per cent worse), demeaning content (25 per cent worse) and condescending responses (15 per cent worse).

•In health-related bias testing, LLMs were from 13 to 17 times more likely to produce a negative story for a character with health conditions than for a healthy one.

2. Why does this happen?

Bias can enter at several stages: training data, data filtering, post-training alignment, reward modeling, system prompts, retrieval sources, product policy, and even the user's own wording.

Major factors cited:

•Training data imbalance: The vast majority of public text online, Wikipedia, news articles, social media, books, comes from sources with a left-leaning tilt. The curation process often favors dominant online voices, which skew toward certain ideologies.

•Data Bias: Chatbots learn from data, and if that data is biased, so are their responses. If historical data favors certain demographics, the chatbot inherits those biases.

•Alignment (RLHF): After pre-training, human annotators guide models toward safe and helpful responses, ranking answers for neutrality and ethical tone. But researchers found reward models showed left-leaning bias even when trained on factual data.

3. Does it matter to users?

Yes, unintentionally.

This unintended power to persuade is caused by latent biases introduced during the training of the large language models that drive chatbots' core capabilities.

In a preregistered experiment with 1,912 participants, participants read Wikipedia summaries or GPT-4o summaries of two historical events. Default AI summaries led to more liberal opinions than Wikipedia summaries, with a reported mean difference of 0.10, Cohen's d = 0.14. The shift came from framing rather than factual errors.

4. Important nuances

•Bias is not a single fixed trait. It varies by model, prompt, topic, language, scale, training method, guardrails, and whether the model is being forced to answer or allowed to refuse.

•Not all models lean the same way. The question is not only whether LLMs as a class lean left; it is which product, under which instructions, gives which kind of answer today. Some models like Grok have been found to lean right, and Gemini was found to give balanced, both-sides answers more often.

•Bias can be reduced, not eliminated. When models were prompted to take a more neutral stance, users perceived less slant.

In short, modern chatbots reflect patterns in their training data and alignment process, and multiple independent methods (human ratings, comparison with party platforms, political-compass tests, and opinion-shift experiments) converge on finding systematic biases — most studied are political left-leaning and gender/racial stereotypes — but the strength and direction depend heavily on the specific model and how you ask the question.

Citations (3)

1etcjournal.comNews and media
AI Chatbots Are Liberal-Left ‘Biased’
Backs: “Yes, the evidence shows AI chatbots can be biased — but "bias" does not mean the models have beliefs or conscious political intentions. Researchers use "bias" to mean a systematic tendency in outputs rather than a conscious political intention, or more formally a systematic pattern in model outputs that favors one…”
2techradar.comCommercial contentquality 0.64
New research shows AI chatbots suggesting women and minorities should be paid less.
Backs: “A UN Women analysis released in June 2026 reviewed 133 AI systems. Forty-four percent showed clear gender bias. A more direct economic test found AI chatbots routinely suggest lower salaries to women and some ethnic minorities and people who described themselves as refugees, even when the job, their qualifications,…”
3archynewsy.comNot yet classified
Chatbots Exhibit Stereotypes & Bias in Representing Demographics, Study Finds
Backs: “BERT, a leading NLP model developed by Google, has exhibited racial biases in various tasks. A Penn State study of GPT-4o, Gemini 1.5 Pro, and DeepSeek v2.5 found LLMs often rely on harmful stereotypes when generating personas based on sociodemographic factors.”