Are AI chatbots biased?

Yes

Yes, AI chatbots are biased—in multiple measurable ways. This is not speculation; it is repeatedly documented in independent tests, academic studies, and real-world evaluations. Bias arises primarily because models are trained on vast internet-scale data reflecting human societies (with their prejudices, imbalances, and cultural skews), then further shaped by fine-tuning, RLHF (reinforcement learning from human feedback), safety guardrails, and company values. They do not have "opinions," but their outputs systematically favor certain framings, groups, or viewpoints over others.[1]

Political Bias (Most Documented)

Numerous studies and media tests find a predominant left-leaning / progressive tilt in major Western models:

•A 2026 Washington Post analysis tested models on hot-button issues (affirmative action, Citizens United, electoral college, taxes, policing, etc.). OpenAI’s ChatGPT (and successors) answered nearly exclusively with left-leaning arguments (~80% only-left, right-leaning almost never). Anthropic’s Claude leaned left or both-sides (rarely pure right). Google’s Gemini was more balanced (both-sides >90%). Even xAI’s Grok (marketed as less “woke”) cited left-leaning arguments more often on average, though it produced more right-leaning responses than peers.[1]

•Academic work (Political Compass tests, Wahl-O-Mat alignments in Germany/Europe, multilingual evaluations, US voter comparisons) consistently shows most conversational LLMs (ChatGPT family, Claude, Llama, Gemini, etc.) cluster left-of-center or libertarian-left. Larger models often show stronger left bias. Base (pre-instruction) models are less coherent but the tuned chat versions amplify the tilt. Biases vary by language, topic polarization, and prompting; they can be steered.[2]

•Models can persuade users: default (often left-tilted) summaries shift opinions leftward relative to neutral sources like Wikipedia; they are highly persuasive (sometimes via inaccurate claims), and can chameleon to the user’s stated politics.[3]

•Exceptions and variations exist: some models refuse political questions more; Grok and certain open models can be less left-skewed or right-shifted via fine-tuning; Chinese models may show different foreign-policy hawkishness. “Neutral” claims by companies often conflict with output patterns.[4]

This matches earlier findings (e.g., Rozado’s multi-test analyses) and holds across US, European, and other contexts, though the specific left/right mapping depends on the political spectrum used.

Other Forms of Bias

•Demographic and social: Ageism (more stereotyping of older people than gender, despite anti-sexism tuning); reduced empathy or accuracy for certain races, mental health conditions (e.g., higher stigma toward schizophrenia/alcohol dependence), lower-education/non-native English/non-US users (worse accuracy, more refusals, condescension); face-based “trustworthiness/competence” biases in hiring-like scenarios that amplify human stereotypes.[5]

•Religious: Under-mention of religion in moral/life questions relative to human expectations; differential positive/negative steering toward certain faiths.[6]

•Other: Mental health advice that can validate delusions or fail crisis detection; source fabrication with asymmetric accuracy on left- vs. right-coded claims in some tests; commercial/self-promotion patterns.

Why Does This Happen?

•Training data: Internet text, books, news, and code over-represent certain demographics, ideologies (e.g., academic/media left-leaning content in English data), and cultural norms.

•Post-training: Human raters (often from specific cultural/ideological pools), safety layers that suppress “harmful” or “controversial” content unevenly, and company goals (e.g., avoiding offense on progressive priorities more aggressively than others).

•Not inevitable or uniform: Biases can be measured, mitigated (better data balancing, diverse raters, explicit even-handedness training, activation steering), or intentionally shifted. Smaller/base models or differently tuned ones (e.g., some open-source or Grok variants) differ. Models also mirror user cues strongly (“sycophancy”).

•No true neutrality: Every design choice (what to refuse, how to frame trade-offs, which sources to weight) embeds values. Claims of pure neutrality are marketing; useful systems reflect some worldview.

Practical Implications

Users should treat chatbots as tools with known skews, not oracles. Cross-check important claims (especially political, medical, legal, or hiring-related), prompt for multiple perspectives (“steelman the opposing view,” “what would a conservative/libertarian/etc. argue?”), compare models, and prefer those with transparent evaluations. Bias can amplify polarization, misinform, or discriminate at scale as adoption grows. Research continues on detection (e.g., Polar benchmark, evenhandedness scores) and reduction.[7]

In short: Yes—systematically, in politics and beyond. The evidence is broad and consistent across independent sources. Awareness + critical use is the best countermeasure.

Citations (7)

1washingtonpost.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: “They do not have "opinions," but their outputs systematically favor certain framings, groups, or viewpoints over others.”
statistic, article, points to yes. Claim: Washington Post test: ChatGPT answered ~80% with only left-leaning arguments, Claude leaned left, Gemini more balanced, Grok still cited left more often.
2aclanthology.orgAcademic
https://aclanthology.org/2025.findings-ijcnlp.17/
Backs: “Biases vary by language, topic polarization, and prompting; they can be steered.”
evidence, article, points to yes. Claim: Academic work shows most conversational LLMs cluster left-of-center; larger and tuned models show stronger tilt, varying by language and prompting.
3futurity.orgNews and media
https://www.futurity.org/ai-chatbots-opinions-3325222/
Backs: “Models can persuade users: default (often left-tilted) summaries shift opinions leftward relative to neutral sources like Wikipedia; they are highly persuasive (sometimes via inaccurate claims), and can chameleon to the user’s stated politics.”
evidence, article, points to yes. Claim: Default left-tilted chatbot summaries shift user opinions leftward, and models are persuasive and mirror users' stated politics.
4brookings.eduNonprofits and think tanks
https://www.brookings.edu/articles/is-the-politicization-of-generative-ai-inevitable/
Backs: ““Neutral” claims by companies often conflict with output patterns.”
evidence, article, points to yes. Claim: Exceptions exist: some models refuse political questions, Grok or open models can be right-shifted; company neutrality claims conflict with outputs.
5uoc.eduAcademic
https://www.uoc.edu/en/news/2026/ageism-ai-chatbots-age-discrimination
Backs: “Ageism (more stereotyping of older people than gender, despite anti-sexism tuning); reduced empathy or accuracy for certain races, mental health conditions (e.g., higher stigma toward schizophrenia/alcohol dependence), lower-education/non-native English/non-US users (worse accuracy, more refusals, condescension);…”
evidence, article, points to yes. Claim: Demographic biases such as ageism, stigma toward mental health conditions, and worse performance for lower-education or non-native English users.
6axios.comNews and mediaquality 0.90
https://www.axios.com/2026/06/01/ai-religious-bias-catholics-chatbots
Backs: “Under-mention of religion in moral/life questions relative to human expectations; differential positive/negative steering toward certain faiths.”
evidence, article, points to yes. Claim: Religious bias: under-mention of religion in moral questions and differential steering toward certain faiths.
7axios.comNews and mediaquality 0.90
https://www.axios.com/2025/11/13/anthropic-bot-bias-data
Backs: “Research continues on detection (e.g., Polar benchmark, evenhandedness scores) and reduction.”
further reading, article, points to yes. Claim: Research continues on bias detection and reduction, such as evenhandedness scores and benchmarks.