Are AI chatbots biased?

YesSays it is contested

Yes, AI chatbots (powered by large language models or LLMs) exhibit various biases. These arise primarily because models are trained on vast amounts of human-generated text and data from the internet, books, and other sources that reflect real-world societal imbalances, stereotypes, historical prejudices, and cultural skews. Biases can also enter or be amplified during fine-tuning, reinforcement learning from human feedback (RLHF/alignment), data filtering, and design choices. While companies work to mitigate overt harms, residual, latent, and implicit biases often remain.[1]

Sources of Bias

•Training data: Web-scale corpora overrepresent certain demographics (e.g., Western, English-speaking, higher-education, Global North sources), occupations, viewpoints, and historical patterns. Biases in the data (e.g., gender-occupation stereotypes) are often correlated with—and sometimes amplified in—model outputs.[2]

•Alignment and post-training: Efforts to make models “helpful, harmless, and honest” or value-aligned can suppress explicit bigotry but leave implicit associations, or introduce new skews based on the preferences of annotators/raters (who may lean certain ways politically or culturally). Instruction-tuning can partially reduce some representational biases while preserving stereotypical associations.[3]

•Other factors: Prompting, decoding hyperparameters, user demographics inferred from language/style, and even position or length biases in responses. Models can detect or infer race/gender/etc. from context and adjust (sometimes unequally).

Bias is not always intentional malice; it often mirrors or exaggerates patterns in society and data. Defining “unbiased” or “neutral” is itself contested, especially on political or value-laden topics.

Types and Evidence of Bias

Social and demographic biases (gender, race, religion, etc.):

•Models frequently reinforce stereotypes (e.g., associating certain genders with occupations more strongly than real-world statistics or even human perceptions; recommending roles differently by name/ethnicity).[4]

•Studies find reduced empathy or less helpful responses for certain racial groups (e.g., Black or Asian users in mental health contexts), discrimination in simulated hiring/CV screening by ethnicity/gender, and unequal treatment.[5]

•Explicitly “unbiased” or value-aligned models still form biased associations on psychology-inspired tests (race/criminality, gender/science, etc.) and can lead to discriminatory decisions.[6]

•Performance can degrade for users with lower English proficiency, less formal education, or non-US origins (less accurate/truthful answers, more refusals, condescending language).[7]

•Gender, racial, and religious biases appear in generated content and narratives across models like ChatGPT variants, Gemini, and others, though newer/larger models sometimes show improvements in consistency.[8]

Political and ideological biases:

•Many prominent models (ChatGPT/GPT series, Claude, Gemini) show left-leaning or liberal tendencies on policy questions, historical framing, and opinion probes—often presenting left-leaning arguments more readily or exclusively, while being more even-handed or resistant on right-leaning ones. Grok (and some others) has been designed or tuned toward less progressive capture and can show different (sometimes right-leaning or more contrarian) patterns, though results vary by test and version.[9]

•Default summaries of historical events can subtly shift user opinions leftward due to latent framing biases, even when factually accurate and unprompted for persuasion. Prompted conservative framing has more limited effects.[10]

•Chatbots can be more persuasive than traditional ads at shifting political views (citing “facts,” though accuracy varies, and some find right-leaning advocacy linked to more inaccuracies in certain studies).[11]

•Biases appear in candidate endorsements, issue stances, and source selection when models search or generate political content.

Other: Cultural/regional skews, age stereotypes, toxicity differences, and amplification of certain narratives. Models can also exhibit “moods,” personalities, or hidden concepts that researchers can probe and steer.[12]

Impacts and Real-World Relevance

Biases can compound with frequent use: influencing opinions on history/politics, providing unequal quality of service, perpetuating stereotypes in recommendations/generation, or affecting high-stakes areas (therapy-like advice, hiring tools, information seeking). They raise fairness, trust, and societal concerns, especially as chatbots become ubiquitous.

Citations (12)

1aclanthology.orgAcademic
https://aclanthology.org/2024.cl-3.8/
Backs: “While companies work to mitigate overt harms, residual, latent, and implicit biases often remain.”
evidence, article, points to yes. Claim: LLM chatbots exhibit biases arising from training data, fine-tuning and design, with residual implicit biases remaining despite mitigation.
2aclanthology.orgAcademic
https://aclanthology.org/2026.lrec-1.315/
Backs: “Biases in the data (e.g., gender-occupation stereotypes) are often correlated with—and sometimes amplified in—model outputs.”
evidence, article, points to yes. Claim: Training data overrepresents certain demographics; data biases like gender-occupation stereotypes correlate with and are sometimes amplified in model outputs.
3pnas.orgAcademic
https://www.pnas.org/doi/10.1073/pnas.2416228122
Backs: “Instruction-tuning can partially reduce some representational biases while preserving stereotypical associations.”
evidence, article, points to yes. Claim: Alignment suppresses explicit bigotry but leaves implicit associations or introduces new skews from annotator preferences.
4acm.orgAcademic
https://dl.acm.org/doi/10.1145/3582269.3615599
Backs: “Models frequently reinforce stereotypes (e.g., associating certain genders with occupations more strongly than real-world statistics or even human perceptions; recommending roles differently by name/ethnicity).”
evidence, article, points to yes. Claim: Models reinforce stereotypes, such as gender-occupation associations and differing role recommendations by name or ethnicity.
5mit.eduAcademic
https://news.mit.edu/2024/study-reveals-ai-chatbots-can-detect-race-but-racial-bias-reduces-response-empathy-1216
Backs: “Studies find reduced empathy or less helpful responses for certain racial groups (e.g., Black or Asian users in mental health contexts), discrimination in simulated hiring/CV screening by ethnicity/gender, and unequal treatment.”
evidence, article, points to yes. Claim: Chatbots show reduced empathy for certain racial groups in mental health contexts and discriminate in simulated hiring.
6pmc.ncbi.nlm.nih.govAcademic
https://pmc.ncbi.nlm.nih.gov/articles/PMC11874501/
Backs: “Explicitly “unbiased” or value-aligned models still form biased associations on psychology-inspired tests (race/criminality, gender/science, etc.) and can lead to discriminatory decisions.”
evidence, article, points to yes. Claim: Value-aligned models still form biased associations on psychology-inspired tests and can lead to discriminatory decisions.
7mit.eduAcademic
https://news.mit.edu/2026/study-ai-chatbots-provide-less-accurate-information-vulnerable-users-0219
Backs: “Performance can degrade for users with lower English proficiency, less formal education, or non-US origins (less accurate/truthful answers, more refusals, condescending language).”
evidence, article, points to yes. Claim: Performance degrades for users with lower English proficiency, less education, or non-US origins.
8emerald.comAcademic
https://www.emerald.com/intr/article/doi/10.1108/INTR-10-2024-1536/1272048/Digital-prejudices-an-analysis-of-gender-racial
Backs: “Gender, racial, and religious biases appear in generated content and narratives across models like ChatGPT variants, Gemini, and others, though newer/larger models sometimes show improvements in consistency.”
evidence, article, points to yes. Claim: Gender, racial, and religious biases appear in content across models, though newer models sometimes improve.
9foxnews.comNews and mediaquality 0.53
https://www.foxnews.com/media/most-prominent-ai-chatbots-have-liberal-bias-new-study-finds
Backs: “Grok (and some others) has been designed or tuned toward less progressive capture and can show different (sometimes right-leaning or more contrarian) patterns, though results vary by test and version.”
evidence, article, points to yes. Claim: Prominent models show left-leaning tendencies, while Grok and others show different patterns that vary by test and version.
10yale.eduAcademic
https://news.yale.edu/2026/03/03/ais-hidden-bias-chatbots-can-influence-opinions-without-trying
Backs: “Default summaries of historical events can subtly shift user opinions leftward due to latent framing biases, even when factually accurate and unprompted for persuasion. Prompted conservative framing has more limited effects.”
evidence, article, points to yes. Claim: Default historical summaries can subtly shift user opinions leftward through latent framing biases.
11technologyreview.comNews and mediaquality 0.91
https://www.technologyreview.com/2025/12/04/1128824/ai-chatbots-can-sway-voters-better-than-political-advertisements/
Backs: “Chatbots can be more persuasive than traditional ads at shifting political views (citing “facts,” though accuracy varies, and some find right-leaning advocacy linked to more inaccuracies in certain studies).”
evidence, article, points to yes. Claim: Chatbots can be more persuasive than ads at shifting political views, with accuracy varying.
12mit.eduAcademic
https://news.mit.edu/2026/exposing-biases-moods-personalities-hidden-large-language-models-0219
Backs: “Models can also exhibit “moods,” personalities, or hidden concepts that researchers can probe and steer.”
evidence, article, points to yes. Claim: Models can exhibit moods, personalities, or hidden concepts that researchers can probe and steer.