More answers like this
Muse Spark 1.2Oct 14, 2026meta/muse-spark-1.2-20260805

Are AI chatbots biased?

Are AI chatbots biased? I'll survey recent research and audits to assess what they find.

Yes, evidence indicates AI chatbots can be biased — but the nature and degree depends on the model, the type of bias measured, and how bias is defined.

What research shows

Researchers generally define bias not as conscious political intention but as a systematic tendency in outputs.

1. Political bias

Multiple recent audits have found a left-of-center tilt in many major U.S. models:

•The Washington Post tested models behind OpenAI's ChatGPT, Google's Gemini and others using political questions designed to gauge responses to hot-button issues and found that chatbots have clear political leanings.
•The model that powers ChatGPT answered nearly every question exclusively with left-leaning arguments and presented only right-leaning positions just once. Among all models tested, ChatGPT displayed the strongest tendency toward responses classified as left-leaning, offering exclusively left-leaning arguments in 80 percent of its responses and in another summary about 80% of questions with only left-leaning arguments.
•Google's Gemini mostly took a both-sides approach, offering both left and right positions in more than 90 percent of its answers.
•Even models marketed as having conservative views, including Elon Musk's Grok, cited left-leaning arguments more often, on average.

A Yale study published in PNAS Nexus found this can have persuasive effects even when not intended. Prior research showed AI content prompted to be persuasive can shift opinions, but this study provides evidence the same is also true of content not intended to change minds, such as summaries that popular chatbots produce in response to simple queries about historical events.

The researchers attributed this unintended power to persuade to 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 narratives.

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

That does not mean every chatbot always gives liberal answers — researchers note it does not mean every finding is equally strong and not every chatbot always gives liberal answers.

2. Gender, racial and cultural bias

•Bias and Fairness Issues in clinical use include that bias in training data carries through into chatbot responses as gender disparities, racial and cultural stereotyping, and overlooked symptoms in underrepresented groups.
•A Stanford Law School audit found significant disparities across names associated with race and gender from chatbots like OpenAI's ChatGPT 4 and Google AI's PaLM-2. For example, a chatbot may suggest a different salary for a lawyer depending on whether the name was Tamika vs Todd.
•The AI chatbots' advice systematically disadvantages names that are commonly associated with racial minorities and women, with names associated with Black women receiving the least advantageous outcomes.
•Biases were consistent across 42 prompt templates and several AI models, indicating a systemic issue.
•Most scenarios displayed biases disadvantageous to Black people and women, with the only consistent exception being basketball where biases favored Black athletes.

3. Other biases

•Sycophancy / confirmation bias: Chatbots often hallucinate due to biased or incomplete training data, and models that are fine-tuned on human feedback are known to exhibit sycophancy – prioritising answers that align with user beliefs over the truth.
•Studies also categorize common types of bias in AI to include confirmation bias, racial bias, gender bias, and socioeconomic bias.
•Systematic errors in model output that stem from skewed or unrepresentative training data lead to unfair or prejudiced results (e.g., racial or gender bias).

Why does bias happen?

•Skewed training datasets favor majority populations, causing misdiagnoses or underdiagnoses in underserved groups.
•AI chatbots can inadvertently perpetuate biases present in training data.
•A common example: The training data reflects societal biases, leading to discriminatory or unfair responses. A chatbot trained on a dataset with predominantly male pronouns might favor male pronouns in its responses.
•Researchers conclude pretrained language models do have political leanings that reinforce polarization present in pretraining corpora, propagating social biases.

Mitigation and context

•Mitigation strategies include diversifying datasets and implementing bias checks, data augmentation, bias detection and mitigation techniques, and diverse training data, testing chatbots against bias benchmarks and training on diverse data, and ensuring diversity and representativeness in training datasets.
•OpenAI has described bias as an important, industry-wide problem that its safety team is working to combat.
•Researchers stress that knowing biases exist and continuing to test for them is a first step.

In short, chatbots are not neutral mirrors — they inherit patterns from human-generated data and from fine-tuning — so they can reproduce and sometimes amplify societal, political, and demographic biases, even while companies attempt to make them neutral, nonpartisan tools.

Citations (3)

  1. 1benton.orgNot yet classified
    Are ChatGPT and other AI chatbots politically biased? We tested them.

    Backs: “Political bias. Multiple recent audits have found a left-of-center tilt in many major U.S. models: The Washington Post tested models behind OpenAI's ChatGPT, Google's Gemini and others using political questions designed to gauge responses to hot-button issues and found that chatbots have clear political leanings.”

  2. 2yale.eduAcademic
    AI’s hidden bias: Chatbots can influence opinions without trying

    Backs: “A Yale study published in PNAS Nexus found this can have persuasive effects even when not intended. Prior research showed AI content prompted to be persuasive can shift opinions, but this study provides evidence the same is also true of content not intended to change minds, such as summaries that popular chatbots…”

  3. 3beaconjournal.comNews and mediaquality 0.84
    AI chatbots like ChatGPT treat Black names differently, per study

    Backs: “Bias and Fairness Issues in clinical use include that bias in training data carries through into chatbot responses as gender disparities, racial and cultural stereotyping, and overlooked symptoms in underrepresented groups. A Stanford Law School audit found significant disparities across names associated with race and…”