Are AI chatbots biased?

I'll pull current research to give you a clear answer on whether AI chatbots are biased.

Yes, research consistently finds that AI chatbots are biased - but the type, degree and cause of bias varies by model, topic and way it is measured.

1. Multiple studies have documented bias

A University of Washington study developed a system to detect subtle biases and found seven of the eight popular AI models they tested in conversations around race and caste generated significant amounts of biased text — particularly when discussing caste.

The researchers generated 1,920 conversations around race (Black and white) and caste (Brahmin and Dalit) in mock hiring contexts. Overall, the team found that 69% of conversations around caste and 48% of conversations overall contained harmful content. For example, a model failed a competence metric when it wrote “our team is mostly White, and he might have trouble communicating with them”.

Both ChatGPT models generated significantly less harmful conversation than the six open source models, but even they were not equivalent: one generated no harmful content about race, but significantly more on caste.

> Systemic biases still can arise subtly in chatbot interactions. Also, since many systems are created in Western countries, their guardrails don’t always recognize non-Western social concepts, such as caste.

To categorize these covert harms, the team drew on social science theories to create the Covert Harms and Social Threats (CHAST) framework, which includes metrics like “competence threats” and “symbolic threats”.

2. Political bias is one of the most studied areas

Multiple prior studies have found that ChatGPT has a left-wing political bias.

A 2025 Stanford study used a more "ecologically valid" method - having more than 10,000 people rate actual responses to 30 political questions from 24 LLMs:

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

Collectively, OpenAI models had the most intensely perceived left-leaning slant – four times greater than perceptions of Google, whose models were perceived as the least slanted overall. On average, models from Google and DeepSeek were seen as statistically indistinguishable from neutral, while models from Elon Musk’s xAI, which touts its commitment to unbiased output, were perceived as exhibiting the second-highest degree of left-leaning slant.

Yale research notes this unintended power to persuade is caused by latent biases introduced during training of the large language models that drive chatbots, which can carry over from ideological leanings in the data used to train LLMs.

3. Other common forms of bias

AI bias is defined as the systematic tendency of a model to produce outputs that are skewed, inequitable, or inaccurate for certain groups, typically due to flawed or unrepresentative data. It enters the pipeline quietly: in how training data is collected, how labels are assigned, how evaluation benchmarks are constructed, and whose feedback is used to fine-tune models.

Documented examples include:

a) Racial and medical bias: A 2025 study in NPJ Digital Medicine tested four leading LLMs (Claude, ChatGPT, Gemini, and NewMes-15) on psychiatric cases. Treatment recommendations changed when race was introduced — two models omitted medication recommendations for ADHD when race was stated explicitly, another suggested guardianship for depression cases with explicit racial identifiers.

b) Gender bias in summarization: Google's Gemma, tested on 29,616 summary pairs from adult social care notes, consistently used more clinical, higher-severity language for male patients. Terms like "disabled," "unable," and "complex medical history" appeared significantly more often in male summaries.

c) Hiring bias: Amazon built a resume-scoring tool trained on ten years of historical applications. Because tech hiring had been heavily male-dominated, the model learned to treat male candidates as the template for success, downgrading resumes containing the word "women's" and graduates of all-women's colleges.

d) Gender bias in recommendations: Training data used to develop LLMs are often biased because they’re based on humanity’s past written records — many of which have historically depicted men as active workers and women as passive objects.

The training data problem is systematic: skewed input data, false logic or just the prejudices of their programmers mean AIs all too easily reproduce and even amplify human biases.

4. Why it matters and can it be fixed?

Studies show biased chatbots can sway users: Democrats and Republicans were both likelier to lean in the direction of the biased chatbot they were talking with than those who interacted with the base model.

However, researchers also found that prompting a model to adopt a neutral stance generated responses that users found less biased and considered to be higher quality. Those neutral responses were more likely to include words like “balance,” “careful,” “complex,” and “sides”.

Experts caution there is no perfect neutral. Hall acknowledges that relying on users’ perceptions to measure political slant has weaknesses, as what users think is neutral may not always be factually correct; what is factually correct may not always be perceived as neutral. By veering toward a neutral stance, however, LLMs could inadvertently reinforce the status quo — which is, in its own way, a kind of slant.

For technical teams, bias is now a legal as well as ethical risk, with the EU AI Act requiring bias detection and mitigation systems and human oversight for high-risk applications, with penalties up to €35 million or 7% of global turnover.

Bottom line: AI chatbots are not neutral mirrors. They reflect and sometimes amplify biases present in training data, evaluation methods, and fine-tuning feedback. Degree of bias varies significantly between models and topics - political lean is most documented as left-libertarian, while racial, gender, caste and dialect biases have been shown to produce concrete harms in hiring, healthcare, and criminal justice contexts. Mitigations include more representative training data, demographically stratified evaluation, and explicit neutral prompting, but no current method removes all bias entirely.