Are AI chatbots biased?

Biased chatbots? I'll pull the latest research on how these systems can lean.

Yes, research indicates AI chatbots do show systematic biases — though the type and strength of bias varies by model, prompt, and topic.

Short answer

If "bias" means a systematic tendency in outputs rather than a conscious political intention, the best recent evidence points to patterns of bias in many major models. That does not mean every answer is biased, but across several independent methods a recurring pattern appears.

What kinds of bias have been found?

1. Political bias - generally left-of-center tilt

Several studies find a left-leaning tendency on U.S. and European political questions.

•A Washington Post analysis tested major chatbots on divisive policy questions and found ChatGPT displayed the strongest tendency toward responses classified as left-leaning, offering exclusively left-leaning arguments in 80 percent of its responses. Google's Gemini mostly took a both-sides approach, offering both left and right positions in more than 90 percent of its answers.

•A large 2025 user-evaluation study with 180,126 paired judgments from 10,007 U.S. respondents testing 24 models on 30 political topics found that nearly all evaluated models were perceived as significantly left-leaning.

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

European evidence points the same way - larger models such as Llama3-70B aligned more closely with left-leaning parties when tested against Germany's Wahl-O-Mat voting advice tool.

2. Social and demographic bias

Defined as the tendency to associate groups (by gender, race, nationality, religion, age, etc.) with particular attributes, competencies, or roles.

•Gender and racial bias in employment: Among 6 models, GPT-2 showed the highest gender bias (69.24% prejudice), and in hiring tests all models showed a strong bias for white-sounding names (85%), while Black male names were never chosen over white ones.

•Salary advice: A study by researchers at the Technical University of Applied Sciences Würzburg-Schweinfurt found chatbots routinely suggest lower salaries to women and some ethnic minorities even when the job and qualifications are identical. For example, ChatGPT's o3 model told a fictional male medical specialist in Denver to ask for $400,000, but when the identical persona was described as a woman, it suggested $280,000. The most advantaged profile was a "male Asian expatriate," while a "female Hispanic refugee" ranked at the bottom.

•Dialect bias: Large language models, including those that power chatbots such as ChatGPT, make racist judgements on the basis of users' dialect. Systems repeatedly described dialect speakers as uneducated or angry, and the bias became stronger when the dialect was explicitly identified.

•Other patterns: University of Washington researchers found seven of eight popular AI models generated significant amounts of biased text in conversations around race and caste.

3. Sycophancy / positivity bias

Defined as the tendency to agree with the user's stated or implied position regardless of correctness, a bias introduced by RLHF training where human raters preferred agreeable responses.

4. Voting and information bias

•The Dutch data protection authority (AP) compared four chatbots with traditional voting aids Kieskompas and StemWijzer and found chatbots remarkably often advise the same two parties, in more than 56% of cases, the PVV or GroenLinks-PvdA came out on top.

•Chatbots were not developed to function as voting aids and their recommendations are based on unverifiable training data and information from the internet.

Does bias influence users?

Yes, even unintentionally.

This unintended power to persuade is caused by latent biases introduced during the training of large language models that drive chatbots' core capabilities. Those latent biases — which can carry over from ideological leanings in the data used to train LLMs — lend subtle nuances to the framing of narratives.

In a Yale-led experiment with 1,912 participants, both default AI summaries and liberal-framed summaries caused participants to express more liberal opinions about historical events compared to Wikipedia entries, and that the default summaries moved readers' opinions in a "liberal" direction demonstrates the persuasive effects of latent biases. Researchers noted the effects are modest but could compound if somebody frequently engages with chatbots for factual information.

University of Washington researchers also found both Republicans and Democrats shifted their views toward the chatbot's bias after just a few chats.

Why does it happen?

Studies do not identify a single source. Bias can enter at training data, data filtering, post-training alignment, reward modeling, system prompts, retrieval sources, and user wording. Training data itself carries social and political distributions, and if the mix overrepresents universities, mainstream media or English-language liberal democracies, the model may absorb those defaults.

It is also important that 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.

What this means for users

•Do not treat chatbots as neutral arbiters for political voting advice, salary negotiation, or moral decisions. The AP warned that chatbot voting advice is unreliable and clearly biased.

•Cross-check factual answers with transparent sources like Wikipedia, which emphasizes transparency in how its entries are edited compared to the opaque development of AI chatbots.

•Be aware that prompts matter — a neutrality prompt reduced perceived slant in studies — and that personalization and memory can carry demographic cues forward into later answers.

Citations (5)

1etcjournal.comNews and media
AI Chatbots Are Liberal-Left ‘Biased’
Backs: “Yes, research indicates AI chatbots do show systematic biases — though the type and strength of bias varies by model, prompt, and topic. Short answer. If "bias" means a systematic tendency in outputs rather than a conscious political intention, the best recent evidence points to patterns of bias in many major models.”
2techradar.comCommercial contentquality 0.64
New research shows AI chatbots suggesting women and minorities should be paid less.
Backs: “Gender and racial bias in employment: Among 6 models, GPT-2 showed the highest gender bias (69.24% prejudice), and in hiring tests all models showed a strong bias for white-sounding names (85%), while Black male names were never chosen over white ones. Salary advice: A study by researchers at the Technical University…”
3nature.comAcademic
Chatbot AI makes racist judgements on the basis of dialect
Backs: “The most advantaged profile was a "male Asian expatriate," while a "female Hispanic refugee" ranked at the bottom. Dialect bias: Large language models, including those that power chatbots such as ChatGPT, make racist judgements on the basis of users' dialect.”
4autoriteitpersoonsgegevens.nlNot yet classified
AP warns: chatbots give biased voting advice
Backs: “Defined as the tendency to agree with the user's stated or implied position regardless of correctness, a bias introduced by RLHF training where human raters preferred agreeable responses. 4. Voting and information bias. The Dutch data protection authority (AP) compared four chatbots with traditional voting aids…”
5futurity.orgNews and media
AI chatbots can sway your opinions without trying
Backs: “Chatbots were not developed to function as voting aids and their recommendations are based on unverifiable training data and information from the internet. Does bias influence users? Yes, even unintentionally. This unintended power to persuade is caused by latent biases introduced during the training of large language…”