Are AI chatbots biased?

Biased chatbots? I'll gather the latest research on how these systems show — and measure — bias.

Yes, AI chatbots can be biased — multiple peer-reviewed studies and investigations have found measurable biases, even when the chatbot is not trying to persuade you.

It does not necessarily mean chatbots are intentionally partisan or prejudiced, but that bias can emerge unintentionally from how they are built.

1. What the research shows

a) Political bias - mostly left-leaning in Western studies

•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.

•In a University of Copenhagen analysis, researchers tested models behind ChatGPT and Gemini using a Danish candidate test and found:

> Overall, all of the tested chatbots place themselves at the center or to the left of center on the political spectrum.
They clustered close to parties such as the Social Democratic Party and The Alternative.

•In another experiment in the same study, the Red–Green Alliance, the Moderates, and Liberal Alliance were recommended disproportionately often, while parties such as the Conservative People's Party, Venstre, and the Denmark Democrats were not suggested as first choice at all by some models.

•An award-winning paper presented at the Association for Computational Linguistics tested 14 large language models and determined that OpenAI's ChatGPT and its GPT-4 model to be the most left-leaning libertarian chatbot, while Meta's LLaMA leaned the furthest right and most authoritarian. The authors concluded: "Our findings reveal that pretrained [language models] do have political leanings that reinforce the polarization present in pretraining corpora, propagating social biases".

•A comprehensive investigation in June 2026 by the Washington Post with researchers from Stanford and Dartmouth found that AI chatbots continue to exhibit a distinct left-leaning political bias when responding to sensitive public policy questions.

•In that test, OpenAI's GPT-5.5 answering nearly every hot-button question with left-leaning arguments and Google's Gemini 3.1 Pro offering both left and right positions in over 90% of its answers. Specifically, GPT-5.5 supplied only the left-leaning side in 80% of its answers, whereas Gemini almost always presented both perspectives.

•Even models marketed as conservative alternatives showed the same pattern. Gab's Arya, marketed with conservative principles, produced left-leaning arguments 12 times more often than right-leaning ones.

b) The bias can influence users without trying

A Yale study published March 3 in PNAS Nexus found that content that is not intended to change minds, such as the summaries that popular chatbots produce in response to simple queries about historical events, can still shift opinions.

The researchers explain:
> This unintended power to persuade is caused by latent biases introduced during the training of the large language models (LLMs) 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 the narratives the chatbots generate

Compared to Wikipedia entries, both the default AI summaries and those prompted to have a liberal framing caused participants to express more liberal opinions. That the default summaries moved readers’ opinions in a “liberal” direction demonstrates the persuasive effects of latent biases.

Political biases can manifest in more subtle ways, for example, in which arguments are emphasized, or which parties are recommended more frequently.

c) Gender and racial bias

•Research has shown that biased chatbot responses occur in approximately 30% of interactions involving sensitive topics like race, gender, or socioeconomic status.

•A 2026 AI Bias Report found 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.

•A biased chatbot may respond differently to users based on their perceived gender, race, or age.

•Early examples are well documented: Microsoft's Tay Chatbot (2016) learned from Twitter interactions and quickly began spouting racist and sexist remarks due to exposure to biased content. Other studies note chatbots repeatedly making sexist jokes and recommending higher-paying jobs to men over women.

Another form of bias is sycophancy — reflecting the user's own leanings: "Even if a chatbot isn't designed to be biased, its answers reflect the biases or leanings of the person asking the questions".

2. Why does it happen?

1.Biased training data: Datasets used for training are often sourced from the internet, where harmful stereotypes and prejudices exist. If the datasets used to train chatbots are skewed or unrepresentative, the AI learning from them may produce biased results.

2.Design choices and feedback loops: Design choices made by developers can unknowingly introduce bias and feedback loops where chatbots learn from user-generated content can reinforce and amplify existing biases.

3.Lack of transparency: Researchers note it is not possible to see why a chatbot recommends a particular party, or which assumptions and data its answers are based on. In contrast to Wikipedia, which emphasizes transparency in how its entries are edited, the development of AI chatbots is opaque.

3. Can bias be mitigated?

Researchers and developers are actively working on this. Common strategies include:

•Using diverse, representative datasets and regularly auditing systems for bias, and implementing fairness algorithms

•More specifically, AI bias can be mitigated by using diverse training data, implementing fairness algorithms, and continuous monitoring

•Systematic engineering through prompt design, data refinement, and integration of bias detection tools — one University of Maryland project reported a 0% bias rate across all tested prompts after such interventions

Tools such as IBM's AI Fairness 360 and other automated bias detection tools are increasingly used to audit algorithms against fairness metrics.

However, researchers caution that completely unbiased AI may be unrealistic: "Given this latest research, the idea of completely unbiased AI seems fantastical. In the end, just like us flawed humans, AI appears destined to land somewhere on the political spectrum" — and until more reliable alternatives exist, users are encouraged to use large language models critically and with caution.

Bottom line: Yes, current AI chatbots do show measurable political, gender and racial biases on average, but the degree and direction vary by model, prompt, and topic. Bias is often subtle and latent rather than overt, and it matters most when people use chatbots for factual information, career advice, news summaries, or voting advice.

If you use chatbots for sensitive decisions, it helps to:

•Ask for multiple perspectives explicitly ("present left and right arguments")

•Cross-check with human-curated sources

•Be aware that the framing of your own prompt can shape the answer.

Citations (3)

1phys.orgNews and mediaquality 0.85
Chatbots show political bias and steer voters toward some parties, analysis finds
Backs: “1. What the research shows. a) Political bias - mostly left-leaning in Western studies. 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.”
2decrypt.coNot yet classified
You’re Not Imagining It: Top AI Chatbots Have Political Biases, Researchers Say
Backs: “In another experiment in the same study, the Red–Green Alliance, the Moderates, and Liberal Alliance were recommended disproportionately often, while parties such as the Conservative People's Party, Venstre, and the Denmark Democrats were not suggested as first choice at all by some models. An award-winning paper…”
3yale.eduAcademic
AI’s hidden bias: Chatbots can influence opinions without trying
Backs: “Gab's Arya, marketed with conservative principles, produced left-leaning arguments 12 times more often than right-leaning ones. b) The bias can influence users without trying. A Yale study published March 3 in PNAS Nexus found that content that is not intended to change minds, such as the summaries that popular…”