Are AI chatbots biased?

I need to gather evidence from reliable sources on whether AI chatbots exhibit biases.I'm reviewing key studies in more detail for a balanced view.Yes, AI chatbots (powered by large language models or LLMs) exhibit various forms of bias, though the degree, type, and visibility vary by model, version, training, and prompting. These are well-documented across academic studies, audits, and independent tests. Bias here means systematic tendencies to favor certain viewpoints, stereotypes, groups, or framings over others—often reflecting (and sometimes amplifying) patterns in training data or human feedback—rather than pure neutrality or pure factuality.[1]

Sources of Bias

LLMs learn from massive internet-scale text corpora that contain societal stereotypes, imbalances, historical inequities, and cultural skews (e.g., overrepresentation of Western, English-language, or majority-group perspectives). This leads to inherited biases.[1]

Further shaping occurs via:

•Fine-tuning and alignment (e.g., supervised fine-tuning and Reinforcement Learning from Human Feedback/RLHF): Human raters’ preferences, demographics, and instructions introduce or reinforce biases. Annotators are often not fully representative, and processes can favor certain political or cultural views (e.g., shifting models toward more liberal-leaning outputs in some cases). RLHF can also introduce algorithmic biases or reward hacking.[2]

•System prompts, safety filters, and company values: These guide outputs toward “helpfulness,” “harmlessness,” or specific ideologies, which can create left- or right-leaning tilts or suppress certain topics.

•Other factors: Data contamination, demographic “leaks” in prompts, and the models’ tendency to pattern-match stereotypes.

Biases can be explicit (overt stereotypes or refusals) or implicit (subtle associations that appear even when models pass standard “unbiased” benchmarks).[3]

Evidence of Specific Biases

Social biases (gender, race, ethnicity, religion, etc.):

•Models often associate occupations with genders stereotypically (e.g., more likely to link certain jobs to men or women beyond real-world statistics) and amplify perceptions.[4]

•Racial biases appear in reduced empathy for certain groups (e.g., GPT-4 showed 2–17% lower empathy for Black and Asian posters in mental health support scenarios compared to White or unspecified), word-level associations in generated content, and disparities in job applicant evaluations or medical contexts.[5]

•Intersectional issues: Underrepresentation or subordinated roles for minoritized race/gender/sexual orientation characters in open-ended narratives; stereotypes across categories.[6]

•Religious and other biases documented in multi-chatbot analyses (ChatGPT variants, Gemini, Bing, etc.). Newer/better-aligned models (e.g., GPT-4 vs. earlier) often show reduced but not eliminated bias.[7]

•Even “explicitly unbiased” models form biased associations on psychology-inspired tests (e.g., race-criminality, gender-science) and make discriminatory recommendations (e.g., names influencing job suggestions).[3]

•Performance disparities: Less accurate or more refusal-prone responses for users with lower English proficiency, less education, or non-US origins.[8]

Political biases:

•Many major models lean left on contested issues (e.g., affirmative action, taxes, healthcare, campaign finance, death penalty). In one structured test of hot-button questions, OpenAI’s model gave exclusively left-leaning arguments ~80% of the time (right-leaning only once); others varied, with Gemini often both-sides, and even “anti-woke” models like Grok still citing left-leaning arguments more often on average (though relatively more right-leaning than peers).[9]

•Other tests show models favoring one party in candidate endorsements or election-related queries, with source selection and framing differences. Accuracy and sourcing can also falter, sometimes relying on state media or producing errors.[10]

•Biases are not uniform—company goals (e.g., Grok’s “truth-seeking”/less “woke” framing) and updates shift them, and “neutrality” itself is contested since middle-ground or both-sides approaches involve value judgments.

Other biases: Preference for LLM-generated content over human (potential anti-human discrimination); toxicity or unequal representation in themes; medical imaging or diagnostic shortcuts using demographics.[11]

Studies consistently find these patterns across models like GPT variants, Claude, Gemini, Llama, Grok, and others, though magnitudes differ (ChatGPT often lower on some social biases than earlier models; Claude sometimes stronger on certain safety metrics).[12]

Mitigation and Limitations

Companies and researchers use techniques like diverse/curated data, debiasing during pre-training/fine-tuning, prompt engineering (e.g., instructing models to consider demographics or both sides), RLHF refinements, constitutional AI, and post-processing filters. Explicit instructions can reduce some empathy or stereotype gaps. Surveys catalog metrics, datasets, and methods (pre-/in-/post-processing).[13]

However:

•Complete elimination is extremely difficult or impossible because language and society contain biases; models reflect training distributions.

•Trade-offs exist (e.g., over-refusal, reduced capability, or new biases from alignment).

•Implicit biases persist; benchmarks can be gamed or incomplete.

•Updates and versions change behavior; adversarial prompting can elicit worse outputs.

•“Bias” definitions vary (statistical parity vs. fairness vs. truth-seeking vs. harm avoidance), and perfect neutrality may not be desirable or achievable on value-laden topics.

Practical Implications

•Users: Be critical—cross-check facts, especially on sensitive/political/medical topics; use multiple models; craft careful prompts; recognize that outputs are probabilistic patterns, not oracles.

•Society: Risks include reinforcing stereotypes, unequal access/quality of service (e.g., mental health, hiring, info for vulnerable users), political persuasion (chatbots can sway opinions effectively, sometimes via inaccurate “facts”), and amplification of inequities.[14]

•Progress is real: Newer models are often safer and less overtly biased than early ones, with ongoing research into detection (e.g., probing hidden concepts) and reduction.

In short, AI chatbots are biased because they are trained on biased human data and shaped by human choices, but awareness, better practices, transparency, and technical advances can substantially reduce harmful effects. No major model is free of them, and claims of perfect neutrality should be viewed skeptically. For the latest on a specific model, check recent independent audits, as the field evolves quickly.

Citations (14)

1mit.eduAcademic
https://direct.mit.edu/coli/article/50/3/1097/121961/Bias-and-Fairness-in-Large-Language-Models-A
Backs: “Bias here means systematic tendencies to favor certain viewpoints, stereotypes, groups, or framings over others—often reflecting (and sometimes amplifying) patterns in training data or human feedback—rather than pure neutrality or pure factuality.”
2springer.comAcademic
https://link.springer.com/article/10.1007/s13347-025-00928-y
Backs: “Annotators are often not fully representative, and processes can favor certain political or cultural views (e.g., shifting models toward more liberal-leaning outputs in some cases).”
3pnas.orgAcademic
https://www.pnas.org/doi/10.1073/pnas.2416228122
Backs: “Biases can be explicit (overt stereotypes or refusals) or implicit (subtle associations that appear even when models pass standard “unbiased” benchmarks).”
4acm.orgAcademic
https://dl.acm.org/doi/10.1145/3582269.3615599
Backs: “Models often associate occupations with genders stereotypically (e.g., more likely to link certain jobs to men or women beyond real-world statistics) and amplify perceptions.”
5mit.eduAcademic
https://news.mit.edu/2024/study-reveals-ai-chatbots-can-detect-race-but-racial-bias-reduces-response-empathy-1216
Backs: “Racial biases appear in reduced empathy for certain groups (e.g., GPT-4 showed 2–17% lower empathy for Black and Asian posters in mental health support scenarios compared to White or unspecified), word-level associations in generated content, and disparities in job applicant evaluations or medical contexts.”
6nature.comAcademic
https://www.nature.com/articles/s41467-025-68004-9
Backs: “Intersectional issues: Underrepresentation or subordinated roles for minoritized race/gender/sexual orientation characters in open-ended narratives; stereotypes across categories.”
7emerald.comAcademic
https://www.emerald.com/intr/article/doi/10.1108/INTR-10-2024-1536/1272048/Digital-prejudices-an-analysis-of-gender-racial
Backs: “Religious and other biases documented in multi-chatbot analyses (ChatGPT variants, Gemini, Bing, etc.). Newer/better-aligned models (e.g., GPT-4 vs. earlier) often show reduced but not eliminated bias.”
8mit.eduAcademic
https://news.mit.edu/2026/study-ai-chatbots-provide-less-accurate-information-vulnerable-users-0219
Backs: “Performance disparities: Less accurate or more refusal-prone responses for users with lower English proficiency, less education, or non-US origins.”
9washingtonpost.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: “In one structured test of hot-button questions, OpenAI’s model gave exclusively left-leaning arguments ~80% of the time (right-leaning only once); others varied, with Gemini often both-sides, and even “anti-woke” models like Grok still citing left-leaning arguments more often on average (though relatively more…”
10straitstimes.comNews and mediaquality 0.75
https://www.straitstimes.com/world/united-states/chatbots-struggle-with-news-accuracy-and-sourcing-ahead-of-us-midterms
Backs: “Other tests show models favoring one party in candidate endorsements or election-related queries, with source selection and framing differences. Accuracy and sourcing can also falter, sometimes relying on state media or producing errors.”
11pubmed.ncbi.nlm.nih.govAcademic
https://pubmed.ncbi.nlm.nih.gov/40729390/
Backs: “Other biases: Preference for LLM-generated content over human (potential anti-human discrimination); toxicity or unequal representation in themes; medical imaging or diagnostic shortcuts using demographics.”
12nature.comAcademic
https://www.nature.com/articles/s41598-024-55686-2.pdf
Backs: “Studies consistently find these patterns across models like GPT variants, Claude, Gemini, Llama, Grok, and others, though magnitudes differ (ChatGPT often lower on some social biases than earlier models; Claude sometimes stronger on certain safety metrics).”
13aclanthology.orgAcademic
https://aclanthology.org/2024.cl-3.8/
Backs: “Companies and researchers use techniques like diverse/curated data, debiasing during pre-training/fine-tuning, prompt engineering (e.g., instructing models to consider demographics or both sides), RLHF refinements, constitutional AI, and post-processing filters. Explicit instructions can reduce some empathy or…”
14technologyreview.comNews and mediaquality 0.91
https://www.technologyreview.com/2025/12/04/1128824/ai-chatbots-can-sway-voters-better-than-political-advertisements/
Backs: “Society: Risks include reinforcing stereotypes, unequal access/quality of service (e.g., mental health, hiring, info for vulnerable users), political persuasion (chatbots can sway opinions effectively, sometimes via inaccurate “facts”), and amplification of inequities.”