Are AI chatbots biased?

Yes, AI chatbots (powered by large language models or LLMs) exhibit various biases, though the type, degree, and manifestation vary by model, training, and context. These are well-documented in peer-reviewed research and audits. Biases arise primarily because models learn statistical patterns from vast human-generated data (internet text, books, etc.) that embed societal stereotypes, imbalances, and viewpoints, then undergo fine-tuning and alignment (e.g., RLHF—reinforcement learning from human feedback) that can introduce or amplify further preferences from trainers and labelers.[1]

Types of Bias and Evidence

•Social and stereotype biases (gender, race, ethnicity, religion, age, etc.): Models often associate groups with stereotypes (e.g., linking certain races to criminality/weapons, genders to careers like science vs. humanities, or names to job suitability). Even models that pass explicit “unbiased” benchmarks form implicit biased associations and can make discriminatory recommendations, such as preferring Caucasian-named candidates for supervisor roles or suggesting different fields of study by gender.[2][3] Studies of popular chatbots (ChatGPT variants, Gemini, Bing/Claude equivalents) using prompt tests found unequal representation, stereotypical themes, and sentiment differences across groups.[4] Audits of hiring-like tasks showed ethnic and gender influences on evaluations, with ethnic effects often stronger.[5] Gender bias appears frequently in leadership descriptions, occupation associations, and medical recommendations.[6]

•Political and ideological leanings: Many models show a left-leaning tendency on contested issues (e.g., framing historical events, policy questions on taxes, affirmative action, crime, or free speech). Default summaries or responses can subtly shift user opinions in a liberal direction compared to neutral sources like Wikipedia, even without explicit persuasion prompts—effects are modest but can compound with repeated use.[7][7] Independent tests (e.g., Washington Post analyses of political questions) found models like those powering ChatGPT often providing exclusively left-leaning arguments far more than right-leaning ones; others (including some marketed as less “woke”) still lean left on average or mix sides unevenly, though balance varies (Gemini often more both-sides; Grok relatively less left-skewed in some tests but not fully neutral).[8][9] Latent framing from training data contributes; deliberate prompting or fine-tuning can shift it further.

•Other biases: Cognitive biases mirroring humans (e.g., preference for negative, social, threat-related, or stereotype-consistent content in transmission/summarization).[10] AI-AI bias (favoring LLM-generated content over human).[11] Sensitivity to user identity markers in queries (race/gender/age cues altering medical, legal, salary, or political advice).[12] Models can also develop novel biases through exploration or feedback loops, and human-AI interaction can amplify biases over time.[13][14]

Surveys and reviews of the literature consistently find that the large majority of evaluated studies detect biases across domains, including medicine.[15][1]

Sources

•Pretraining data: Web-scale corpora over-/under-represent groups, cultures, viewpoints, and eras; they contain stereotypes, toxicity, and historical inequities. Western-centric or English-heavy data skews cultural norms.[16]

•Alignment/RLHF and fine-tuning: Human preferences (from raters who may not be demographically or ideologically representative) shape “helpful/harmless” behavior. This can inject collective or researcher biases, make models better at sounding correct (even if wrong), or over-refuse/over-correct in ways that create new imbalances. Demographics of raters and instruction framing matter.[17][18]

•Other: Model architecture, decoding, safety guardrails (which can be uneven), and post-training interventions. Explicit debiasing helps on benchmarks but often leaves residual or implicit biases.[2]

Impacts and Caveats

Biases can lead to unfair outcomes in high-stakes uses (hiring, medicine, advice), reinforce stereotypes, subtly persuade users, or reduce trust. They are not always “anti-truth” (some statistical associations reflect real-world base rates) and “neutrality” is itself contested—equal airtime for all views can include falsehoods. Different models vary: newer/larger or specifically aligned ones (e.g., some GPT, Claude, Gemini, Grok variants) often show reduced overt bias on standard tests but retain subtler ones.[4][19]

Companies mitigate via data filtering, diverse raters, adversarial testing, constitutional AI-style principles, and ongoing evaluations, but complete elimination is extremely difficult (and zero social association may not always be desirable or possible). Users should treat outputs critically, cross-check facts, use varied prompts/models, and avoid over-reliance for consequential decisions.

In short, bias is a real, multi-faceted feature of current AI chatbots rooted in their data-driven nature, not a rare bug. Progress continues, but awareness and scrutiny remain essential.

Citations (19)

1arxiv.orgAcademic
https://arxiv.org/html/2309.00770v3
Backs: “Biases arise primarily because models learn statistical patterns from vast human-generated data (internet text, books, etc.) that embed societal stereotypes, imbalances, and viewpoints, then undergo fine-tuning and alignment (e.g., RLHF—reinforcement learning from human feedback) that can introduce or amplify further…”
2pmc.ncbi.nlm.nih.govAcademic
https://pmc.ncbi.nlm.nih.gov/articles/PMC11874501/
Backs: “Even models that pass explicit “unbiased” benchmarks form implicit biased associations and can make discriminatory recommendations, such as preferring Caucasian-named candidates for supervisor roles or suggesting different fields of study by gender.”
3pnas.orgAcademic
https://www.pnas.org/doi/10.1073/pnas.2416228122
Backs: “Even models that pass explicit “unbiased” benchmarks form implicit biased associations and can make discriminatory recommendations, such as preferring Caucasian-named candidates for supervisor roles or suggesting different fields of study by gender.”
4emerald.comAcademic
https://www.emerald.com/intr/article/doi/10.1108/INTR-10-2024-1536/1272048/Digital-prejudices-an-analysis-of-gender-racial
Backs: “Even models that pass explicit “unbiased” benchmarks form implicit biased associations and can make discriminatory recommendations, such as preferring Caucasian-named candidates for supervisor roles or suggesting different fields of study by gender.Studies of popular chatbots (ChatGPT variants, Gemini, Bing/Claude…”
5sciencedirect.comAcademic
https://www.sciencedirect.com/science/article/pii/S2949882124000148
Backs: “Even models that pass explicit “unbiased” benchmarks form implicit biased associations and can make discriminatory recommendations, such as preferring Caucasian-named candidates for supervisor roles or suggesting different fields of study by gender.Studies of popular chatbots (ChatGPT variants, Gemini, Bing/Claude…”
6researchgate.netAcademic
https://www.researchgate.net/publication/390145411_Gender_Biases_within_Artificial_Intelligence_and_ChatGPT_Evidence_Sources_of_Biases_and_Solutions
Backs: “Even models that pass explicit “unbiased” benchmarks form implicit biased associations and can make discriminatory recommendations, such as preferring Caucasian-named candidates for supervisor roles or suggesting different fields of study by gender.Studies of popular chatbots (ChatGPT variants, Gemini, Bing/Claude…”
7yale.eduAcademic
https://news.yale.edu/2026/03/03/ais-hidden-bias-chatbots-can-influence-opinions-without-trying
Backs: “Default summaries or responses can subtly shift user opinions in a liberal direction compared to neutral sources like Wikipedia, even without explicit persuasion prompts—effects are modest but can compound with repeated use.”
8washingtonpost.comNews and mediaquality 0.82
https://www.washingtonpost.com/technology/interactive/2026/06/24/are-ai-chatbots-like-chatgpt-politically-biased-we-tested-them/
Backs: “Grok relatively less left-skewed in some tests but not fully neutral).”
9the-decoder.comNews and media
https://the-decoder.com/most-major-ai-chatbots-still-lean-left-on-political-questions-even-anti-woke-models-are-no-exception/
Backs: “Grok relatively less left-skewed in some tests but not fully neutral).”
10pnas.orgAcademic
https://www.pnas.org/doi/10.1073/pnas.2313790120
Backs: “Cognitive biases mirroring humans (e.g., preference for negative, social, threat-related, or stereotype-consistent content in transmission/summarization).”
11pnas.orgAcademic
https://www.pnas.org/doi/10.1073/pnas.2415697122
Backs: “Cognitive biases mirroring humans (e.g., preference for negative, social, threat-related, or stereotype-consistent content in transmission/summarization).AI-AI bias (favoring LLM-generated content over human).”
12arxiv.orgAcademic
https://arxiv.org/html/2507.14238v1
Backs: “Cognitive biases mirroring humans (e.g., preference for negative, social, threat-related, or stereotype-consistent content in transmission/summarization).AI-AI bias (favoring LLM-generated content over human).Sensitivity to user identity markers in queries (race/gender/age cues altering medical, legal, salary, or…”
13arxiv.orgAcademic
https://arxiv.org/html/2511.06148v1
Backs: “Cognitive biases mirroring humans (e.g., preference for negative, social, threat-related, or stereotype-consistent content in transmission/summarization).AI-AI bias (favoring LLM-generated content over human).Sensitivity to user identity markers in queries (race/gender/age cues altering medical, legal, salary, or…”
14nature.comAcademic
https://nature.com/articles/s41562-024-02077-2
Backs: “Cognitive biases mirroring humans (e.g., preference for negative, social, threat-related, or stereotype-consistent content in transmission/summarization).AI-AI bias (favoring LLM-generated content over human).Sensitivity to user identity markers in queries (race/gender/age cues altering medical, legal, salary, or…”
15d-nb.infoNot yet classified
https://d-nb.info/1365657868/34
Backs: “Surveys and reviews of the literature consistently find that the large majority of evaluated studies detect biases across domains, including medicine.”
16mdpi.comAcademic
https://www.mdpi.com/2079-9292/15/9/1824
Backs: “Web-scale corpora over-/under-represent groups, cultures, viewpoints, and eras; they contain stereotypes, toxicity, and historical inequities. Western-centric or English-heavy data skews cultural norms.”
17arxiv.orgAcademic
https://arxiv.org/html/2307.15217
Backs: “This can inject collective or researcher biases, make models better at sounding correct (even if wrong), or over-refuse/over-correct in ways that create new imbalances. Demographics of raters and instruction framing matter.”
18springer.comAcademic
https://link.springer.com/article/10.1007/s13347-025-00928-y
Backs: “This can inject collective or researcher biases, make models better at sounding correct (even if wrong), or over-refuse/over-correct in ways that create new imbalances. Demographics of raters and instruction framing matter.”
19arxiv.orgAcademic
https://arxiv.org/html/2510.12857v1
Backs: “Different models vary: newer/larger or specifically aligned ones (e.g., some GPT, Claude, Gemini, Grok variants) often show reduced overt bias on standard tests but retain subtler ones.”