Are AI chatbots biased?

Are chatbots biased? I'll sort through recent studies to see what the evidence shows.

Yes, AI chatbots can be biased - but the answer is more nuanced than a simple yes/no.

Short answer

AI chatbots are biased if "bias" means a systematic tendency in outputs rather than conscious political intention. The best recent evidence points to a left-of-center tilt in many major models on U.S. and European political questions. That does not mean every chatbot always gives liberal answers, or that every finding is equally strong. Bias varies by model, prompt, topic, language, training method, and guardrails.

1. What "bias" means in this context

Researchers usually mean an empirical pattern: a consistent framing that favors one political, moral, or social view over plausible alternatives. A chatbot does not "believe" in liberalism or conservatism. It generates text from training data, alignment procedures, system instructions, retrieval results, and user prompts. The question is whether those processes produce a consistent tilt.

Studies usually measure it in four ways: asking people to rate whether answers sound left or right, comparing answers to party platforms, using political-compass surveys, and testing whether exposure to AI text changes human opinions.

2. Evidence of political bias

Multiple independent studies in 2024-2026 converge on a similar finding:

•A widely cited 2024 paper, “More human than human: measuring ChatGPT political bias," tested ChatGPT's answers to political quizzes hundreds of times and found "robust evidence" of a systematic tilt toward Democrats in the U.S., Lula in Brazil and Labour in the UK.

•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. 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. It reported ChatGPT gave exclusively left-leaning answers 80 percent of the time and right-leaning answers just 3 percent of the time, while Google Gemini usually presented both sides in more than 90 percent of its answers.

OpenAI itself acknowledged the issue in 2025, finding its models are near-objective on neutral prompts but show "moderate" bias on emotionally charged political questions, and that newer GPT-5 models reduce measured bias by about 30% compared to GPT-4o.

Bias is not just in final answers. A 2024 MIT analysis of language reward models — the systems used to align models during training — found they often exhibit left-leaning political preferences, and that optimizing them can increase that bias over time.

And bias is not limited to the U.S. A 2025 study of ChatGPT across countries found ChatGPT's answers partially adapt to local attitudes, but are liberally biased from 35-60% depending on language and country.

3. Evidence of gender, racial and other biases

Political lean is only one type:

•Racial bias in language models: Microsoft's chatbot Tay, designed to engage with the public, started tweeting offensive and racist content, referencing genocide and the Holocaust within 24 hours after learning from Twitter.

•Gender bias in recruitment: Amazon's recruiting AI was trained on CVs of existing executives which heavily skewed toward men. Consequently, the algorithm learned that men were more likely to become top executives and rejected CVs that mentioned women or women-related terms.

•Amplified stereotypes: In 2023, Bloomberg's analysis of 5,000 images found a generative AI model amplified racial and gender stereotypes well beyond their representation in the actual workforce.

•Health-related bias: When researchers asked models including ChatGPT, Grok and Claude directly using standard stigma questionnaires, they actually scored better than humans, showing less obvious bias. However, in contextual judgment tasks they were from 13 to 17 times more likely to produce a negative story for a character with health conditions than for a healthy one.

•Status quo bias: A study of 11 large language models with nearly 55,000 prompts found AI chatbots backed existing plans about twice as often as identical new ones, a status quo bias that can work against climate choices.

•Cognitive bias that changes user behavior: Researchers at UC San Diego found that LLM-generated summaries changed the sentiments of the reviews they summarized in 26.5% of cases, and customers were 32% more likely to buy a product after reading a review summary generated by a chatbot than after reading the original review.

4. Why does it happen?

•Training data bias: GenAI chatbots are trained on large datasets that often contain biases from the real world, typically scraped from the internet, which means they are likely to reflect historical, social, and cultural biases.

•Over-representation: AI chatbots are trained on vast amounts of data created by humans that reflects biases, where data might overrepresent certain demographics or viewpoints while underrepresenting others.

•How patterns are learned: If the word she appears with nurse ten times more often than with engineer, the model is likely to complete “The senior engineer said that ___” with he. AI systems that learn patterns in past data will often be more biased than humans because a pattern of discrimination is easy to copy.

Other sources include data filtering, post-training alignment, reward modeling, system prompts, product policy, and even user wording. In short, a model can be biased without a developer trying to make it biased.

5. Does bias actually influence users?

Yes, recent experiments suggest it does:

•A short interaction with a chatbot can meaningfully shift a voter's opinion about a presidential candidate or proposed policy, moving opposition voters' preferences by 10 percentage points or more in many cases.

•A Yale-led experiment with 1,912 participants found default GPT-4o historical summaries, while factually accurate, moved readers toward more liberal opinions than Wikipedia summaries. Researchers warn there is no truly neutral AI, and that users tend to absorb a model's bias after just a few interactions.

6. Is it the same for every chatbot and can it be reduced?

No. Model differences are large. In the Washington Post test, ChatGPT was far more one-sided than Gemini which mostly presented both sides. PoliticsBench found Grok as the right-leaning exception among eight models.

Bias can be reduced but not necessarily eliminated:

•A neutrality prompt reduced perceived slant in the large 2025 user study.

•OpenAI reports iterative improvements reducing bias in newer models.

•Researchers recommend: diverse and representative training data, regular bias audits using fairness metrics, adversarial de-biasing, counterfactual testing, transparency, and human oversight.

For users: use specific, balanced prompts, ask the model to present multiple perspectives, compare answers across different chatbots, and be aware that factual accuracy does not guarantee neutral framing.

Citations (3)

1etcjournal.comNews and media
AI Chatbots Are Liberal-Left ‘Biased’
Backs: “Yes, AI chatbots can be biased - but the answer is more nuanced than a simple yes/no. Short answer. AI chatbots are biased if "bias" means a systematic tendency in outputs rather than conscious political intention.”
2ts2.techNot yet classified
ChatGPT and the Risks of Deepening Political Polarization and Divides in the AI Era
Backs: “Evidence of political bias. Multiple independent studies in 2024-2026 converge on a similar finding: A widely cited 2024 paper, “More human than human: measuring ChatGPT political bias," tested ChatGPT's answers to political quizzes hundreds of times and found "robust evidence" of a systematic tilt toward Democrats in…”
3toolify.aiNot yet classified
Uncovering the Dark Side of AI: Bias and Discrimination
Backs: “3. Evidence of gender, racial and other biases. Political lean is only one type: Racial bias in language models: Microsoft's chatbot Tay, designed to engage with the public, started tweeting offensive and racist content, referencing genocide and the Holocaust within 24 hours after learning from Twitter.”