Are AI chatbots biased?

I'll research whether AI chatbots show bias and what the evidence reveals.

Yes, AI chatbots can be biased - if "bias" means a systematic tendency in outputs rather than a conscious political intention. The best recent evidence points to biases on several dimensions:

1. Political bias - a recurring left-of-center tilt

Multiple independent studies using different methods have found a left-leaning pattern on U.S. and European political questions.

> AI chatbots are biased, if "bias" means a systematic tendency in outputs rather than a conscious political intention. The best recent evidence points to a left-of-center tilt in many major models

> when models are asked about contested political or social issues, their default framing often looks more liberal, progressive, or center-left than conservative

Key findings:

•Large human evaluation: A 2025 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

•Washington Post test (June 2026): The Post tested models behind ChatGPT, Gemini and others. The model that powers ChatGPT answered nearly every question exclusively with left-leaning arguments and presented only right-leaning positions just once. Google's Gemini mostly took a both-sides approach, offering both left and right positions in more than 90 percent of its answers. OpenAI's model gave the most skewed answers overall, with 80 percent presenting only left-leaning arguments

•University of Copenhagen (2026): Researchers found popular AI chatbots such as ChatGPT and Gemini are not neutral and tend to favor certain political parties when asked who users should vote for. Overall, all of the tested chatbots place themselves at the center or to the left of center on the political spectrum

•Other benchmarks: A March 2026 benchmark of eight prominent models found that seven leaned left and one, Grok, leaned right. A 2026 PoliticsBench and German Wahl-O-Mat studies found similar patterns.

This does not mean every chatbot always gives liberal answers, or that every finding is equally strong. The finding is strongest for user-perceived slant and for benchmark studies that compare model answers with party platforms. Bias varies by model, prompt, topic, language, scale, training method, guardrails, and whether the model is being forced to answer or allowed to refuse.

2. Social biases - gender, race, and stereotypes

•Gender bias in LLMs: Among 6 models, GPT-2 showed the highest gender bias (69.24% prejudice)

•Racial bias in LLMs: In hiring tests, all models showed a strong bias for white-sounding names (85%), while Black male names were never chosen over white ones

•UN Women analysis (June 2026): Reviewed 133 AI systems — Forty-four percent showed clear gender bias. Twenty-six percent combined gender and racial prejudice

•Financial and hiring advice: A new study found AI chatbots often suggest significantly lower salaries to women and minorities and researchers found that AI chatbots often provide recommendations that are inconsistent across platforms — and that the advice can vary by the gender and race of the person asking

•Stereotyping personas: Chatbots often appear human-like, they overemphasize racial markers and flatten complex identities into stereotypes

3. Other forms of bias

•Sycophancy bias: Chatbots affirmed human users at substantially higher rates than humans did, which can reinforce harmful beliefs and make users more convinced of their own rightness.

•Unintended persuasion: Even summaries not intended to change minds can sway opinions. Yale researchers found this unintended power to persuade is caused by latent biases introduced during the training

•Ideological leanings in training data: Those latent biases — which can carry over from ideological leanings in the data used to train LLMs — lend subtle nuances to the framing

4. Why does bias happen?

Bias can enter at several stages: training data, data filtering, post-training alignment, reward modeling, system prompts, retrieval sources, and user wording.

•Internet-scale training data contains societal biases: There have been instances where chatbots exhibited racial or gender biases... These biases stem from the fact that the datasets used for training are often sourced from the internet, where harmful stereotypes and prejudices exist

•Learning patterns, not just facts: AI chatbots learn from large amounts of text found online... Also, if the information it learned includes biased views, the chatbot might repeat them

•Alignment and institutional knowledge: A model trained to sound like mainstream institutional knowledge may therefore look left-leaning even without being trained to favor a party. The technical question is what the outputs do across many prompts and users

5. Does bias matter?

Yes - research shows it can influence users, even when content is factually accurate.

A preregistered Yale-led experiment (N=1,912) found default GPT-4o historical summaries shifted opinion in a liberal direction compared with Wikipedia summaries, with framing rather than false facts driving the shift.

Other multi-university studies found chatting with a politically biased AI model was more effective than political advertisements at nudging voters.

6. Can bias be reduced?

Bias is not fixed. When models were prompted to take a more neutral stance, users perceived less slant. However:

•Different products use different strategies. Gemini and Claude repeatedly refused some political-quiz questions, while others answered every question

•"Neutrality" itself is hard to define in politics, and even "neutral" or both-sides positions are positions themselves

•Companies say they train to reduce bias: Gemini is designed to provide balanced responses that don't favor any political ideology and Anthropic says We train Claude to treat different political viewpoints equally. OpenAI says it works to measure and reduce political bias.

Bottom line: Independent research converges on the finding that many current chatbots show measurable systematic biases - most consistently a modest left-of-center tilt on contested political and cultural issues in the U.S. and Europe, plus persistent gender and racial biases inherited from training data. The degree varies greatly by model, topic, language and how you ask. Transparency about training data, alignment methods and system instructions remains limited, which researchers say makes critical assessment difficult. As a user, treating chatbots as informative but imperfect tools - cross-checking with diverse sources and using neutral prompts - reduces exposure to bias.

Citations (3)

1etcjournal.comNews and media
AI Chatbots Are Liberal-Left ‘Biased’
Backs: “AI chatbots are biased, if "bias" means a systematic tendency in outputs rather than a conscious political intention. The best recent evidence points to a left-of-center tilt in many major models”
2archive.orgNonprofits and think tanks
Are AI chatbots like ChatGPT politically biased? We tested them. - Washington Post
Backs: “Washington Post test (June 2026): The Post tested models behind ChatGPT, Gemini and others. The model that powers ChatGPT answered nearly every question exclusively with left-leaning arguments and presented only right-leaning positions just once.”
3phys.orgNews and mediaquality 0.85
Chatbots show political bias and steer voters toward some parties, analysis finds
Backs: “OpenAI's model gave the most skewed answers overall, with 80 percent presenting only left-leaning arguments. University of Copenhagen (2026): Researchers found popular AI chatbots such as ChatGPT and Gemini are not neutral and tend to favor certain political parties when asked who users should vote for.”