Are AI chatbots biased?

Yes

Are AI chatbots biased? I'll pull the latest research on how bias shows up in AI chat.

Yes, research consistently finds that AI chatbots do exhibit bias — though the type, degree and cause varies by model.

1. Short Answer: Yes, but it's often subtle and unintentional

Modern chatbots based on large language models (LLMs) are trained on massive amounts of human-created text from the internet, books, and other sources. They can therefore absorb and reproduce human biases.

Researchers categorize bias sources in chatbots as coming from data, user interaction, the algorithm itself, etc.. Because models lack transparency and interpretability, mitigating that bias is challenging.

Many prominent LLMs have built-in guards to catch overt biases such as slurs, but systemic biases still can arise subtly in interactions.

That latent bias can influence users without trying - the unintended power to persuade is caused by latent biases introduced during training that carry over from ideological leanings in the data.

2. Evidence for bias

A. Political bias - consistently left-leaning

Multiple independent studies have found a left-leaning or liberal bias, especially in ChatGPT:

•Prior studies have found that ChatGPT 3.5 has a left-wing political bias.

•Further confirmation came from research showing that ChatGPT favors left-leaning views, and an analysis using the Dutch 2021 election questionnaire identified a discernible left-leaning and liberal bias.

•Another publication made ChatGPT take 15 different political affiliation tests, with 14 out of 15 resulting in a left-leaning (i.e., progressive) bias.

•Based on the Political Compass questionnaire, researchers documented robust evidence that ChatGPT presents a significant and sizeable political bias towards the left side.

A 2025 Stanford study tested 24 LLMs from 8 companies with more than 10,000 U.S. users:

•Users overwhelmingly perceive that some of the most popular LLMs have left-leaning political slants.

•For 18 of the 30 political questions, users perceived nearly all of the LLMs’ responses as left-leaning. This was true for both self-identified Republican and Democratic respondents.

•Aggregated by company, OpenAI models had the most intensely perceived left-leaning slant – four times greater than perceptions of Google. Models from Google and DeepSeek were seen as statistically indistinguishable from neutral, while models from Elon Musk's xAI were perceived as exhibiting the second-highest degree of left-leaning slant.

Bias can have real consequences: In experiments, Democrats and Republicans were both more likely to lean in the direction of the biased chatbot they talked with than those who interacted with the base model, and research shows that cognitive biases introduced by LLMs have real consequences on users’ decision making.

B. Race, caste, gender and other social biases

•A University of Washington system for detecting subtle bias tested eight popular AI models in conversations around race and caste and found seven of the eight generated significant amounts of biased text — particularly when discussing caste.

•That study found 69% of conversations around caste and 48% of conversations overall contained harmful content, including competence threats and disparagement. Open-source models fared far worse than two proprietary ChatGPT models.

•An MIT Schwarzman College study found that GPT-4's response empathy levels were reduced for Black (2 to 15 percent lower) and Asian posters (5 to 17 percent lower) compared to white posters.

•Other reviews note that AI-generated content produced by LLMs exhibits substantial gender and racial biases, with ChatGPT tending to exhibit implicit gender bias while other models showed explicit bias.

Related work has also documented biases in:

•Hiring - e.g., ranking job applicants by name perceived race/gender, or penalizing resumes with disability markers

•Healthcare and education applications

•Stereotyping when asked to take on personas based on age, gender, race, occupation

3. Why does it happen?

1.Training data bias: LLMs are trained on internet-crawled material (about 60% in GPT-3), curated content (22%), books (16%) and Wikipedia (3%), which reflects existing human and cultural biases.

2.Western-centric guardrails: Tools that catch harmful responses do very well when harms are overt and common in a Western context — like a racial slur — but may not recognize non-Western social concepts such as caste.

3.Reinforcement Learning from Human Feedback (RLHF): The fine-tuning process that makes models more helpful can also reinforce political leanings - left-leaning models became more left-leaning, and right-leaning ones more right-leaning after further training.

4. Can bias be reduced?

Research suggests yes, partially:

•Stanford researchers found that prompting a model to adopt a neutral stance generated responses that users found less biased and higher quality, more likely to include words like "balance," "careful," "complex," "sides".

•Users were more likely to trust more neutrally worded responses.

•However, neutrality is not a panacea — what users think is neutral may not be factually correct, and veering toward neutrality could inadvertently reinforce the status quo, which is itself a kind of slant.

Some propose offering models tuned to different political viewpoints to give users choice, though that risks creating echo chambers.

5. What to do as a user

•Treat chatbots as assistants that "can make mistakes" and have noticeable partisan bias, not authoritative sources.

•Cross-check important information, especially on political, medical, or hiring decisions, with primary sources.

•Be aware that prompts influence output, and higher knowledge about how AI works appears to reduce susceptibility to persuasion.

•If you need balanced perspectives, explicitly ask the model to "present balanced arguments from multiple sides with evidence and uncertainty acknowledged" — which researchers found reduces perceived slant.

> In sum: AI chatbots are not intentionally partisan or prejudiced in a human sense, but extensive evidence shows they do reproduce measurable political (primarily left-libertarian), racial, gender, and cultural biases inherited from their training data and fine-tuning. The degree varies significantly by model and company.

Citations (3)

1arxiv.orgAcademic
## BIAS AND FAIRNESS IN CHATBOTS: AN OVERVIEW
Backs: “Modern chatbots based on large language models (LLMs) are trained on massive amounts of human-created text from the internet, books, and other sources.”
evidence, official document, points to yes. Claim: LLMs trained on massive human-created text can absorb and reproduce human biases; bias sources include data, user interaction, and algorithm.
2washington.eduAcademic
In the ‘Wild West’ of AI chatbots, subtle biases related to race and caste often go unchecked | UW News
Backs: “Because models lack transparency and interpretability, mitigating that bias is challenging. Many prominent LLMs have built-in guards to catch overt biases such as slurs, but systemic biases still can arise subtly in interactions.”
evidence, article, points to yes. Claim: Mitigating bias is hard given lack of transparency; guardrails catch overt bias but subtle bias, e.g. around race and caste, persists.
3stanford.eduAcademic
Study finds perceived political bias in popular AI models
Backs: “A 2025 Stanford study tested 24 LLMs from 8 companies with more than 10,000 U.S. users: Users overwhelmingly perceive that some of the most popular LLMs have left-leaning political slants.”
statistic, article, points to yes. Claim: Stanford study of 24 LLMs with over 10,000 users found perceived left-leaning slant, strongest for OpenAI models.