Are AI chatbots biased?

Yes

Are AI chatbots showing bias? I'll check the latest research on their political and social leanings.

Yes, research consistently finds AI chatbots can be biased — politically, racially, and by gender — although the type and degree of bias varies by model and task.

1. Political Bias

Several independent studies have found popular chatbots are not neutral when asked political questions:

•Researchers from the University of Copenhagen found that "Popular AI chatbots such as ChatGPT and Gemini are not neutral and tend to favor certain political parties when asked who users should vote for".

•In that Danish election context, "all of the tested chatbots place themselves at the center or to the left of center on the political spectrum". When asked to recommend parties to fictitious voters, "the Red–Green Alliance, the Moderates, and Liberal Alliance were recommended disproportionately often, while parties such as the Conservative People's Party, Venstre (the Liberal Party of Denmark), and the Denmark Democrats were not suggested as first choice at all by some models."

•A Washington Post investigation with researchers from Stanford and Dartmouth found "several leading AI chatbots display a measurable political tilt in their default settings, with OpenAI's ChatGPT producing the strongest left-leaning pattern". The testing suggested ChatGPT "advanced left-leaning arguments in the overwhelming majority of responses" while "Gemini was described as presenting both sides of the debate in more than 90% of its answers".

•A broader academic audit supports this: computer scientist David Rozado from Otago Polytechnic tested 24 different large language models, including models from OpenAI and Google's Gemini, using 11 standard political orientation tests, and found "a consistent left-leaning bias across the models". The researchers concluded that "pretrained [language models] do have political leanings that reinforce the polarization present in pretraining corpora, propagating social biases".

Other research noted that even models marketed as conservative alternatives still leaned left more often than right in controlled tests.

Researchers warn this matters because chatbots can shape opinions even without intent. A Yale study found that "subtle framing choices can nudge readers in a liberal direction without any explicit political prompting".

2. Racial, Gender and Other Social Biases

Beyond politics, studies find bias in sensitive domains:

•A study in the Journal of Financial Planning found guidance could be "biased advice based on race or gender".

•One overview found "biased chatbot responses occur in approximately 30% of interactions involving sensitive topics like race, gender, or socioeconomic status".

•Many chatbots "inadvertently perpetuate or amplify societal biases present in their training data".

A detailed study published in Nature (September 2024) by Valentin Hofmann and colleagues at Oxford illustrated hidden racial bias:

•When asked directly what it thinks about Black people, ChatGPT used words like "ambitious" and "intelligent," but "what if you ask those AI tools what they think about people who use African American English? Now the AI models tend to use words like 'suspicious,' 'aggressive' and 'ignorant.'"

•For obvious racism, most models used favorable words to describe Black people, with ChatGPT using words with an average rating of 1.3. For hidden bias tested via dialect, "The words that models chose to describe speakers of African American English were overwhelmingly negative. ChatGPT's words scored an average of -1.2".

•Real-world impact tests: "AI models gave the death sentence to the person using Standard American English about 23 percent of the time. The person using African American English got sentenced to death roughly 28 percent of the time".

•For jobs, "The models largely assigned users of African American English to 'low status' jobs. Examples included cook or soldier. Those who had tweeted in Standard American English were assigned to 'higher status' jobs. Professor, for example. Or psychologist. Or economist."

The authors noted that "Hidden biases show up even in AI models released in the last few years... had been trained by people deliberately trying to scrub racism", and that human feedback tuning weakened overt stereotypes but left "hidden, or covert, racism virtually unchanged".

Other documented examples include Google pausing Gemini's image generation after it "tended to favor the generation of images of non-white people and was often incorrectly depicted with the wrong race and/or gender", and historical cases like Microsoft's Tay chatbot spouting racist remarks and Amazon's recruiting tool penalizing women's resumes.

3. Why Does This Happen?

Bias is generally not intentional, but stems from how models are built:

•" latent biases introduced during the training of the large language models (LLMs) that drive chatbots' core capabilities... can carry over from ideological leanings in the data used to train LLMs — lend subtle nuances to the framing of the narratives the chatbots generate".

•"If the datasets used to train chatbots are skewed or unrepresentative, the AI learning from them may produce biased results".

•Many datasets are "sourced from the internet, where harmful stereotypes and prejudices exist. Consequently, even though the developers may not intend to create biased systems, the outcomes suggest otherwise".

•Large models are "trained on vast collections of online text, much of it drawn from mainstream media, academic material and legal sources, which can create an initial skew" and later "safety training, designed to reduce hate speech, discrimination and conspiracy theories, may also over-correct".

4. Does This Mean All Answers Are Useless?

Not necessarily:

•Bias tends to be subtle — "which arguments are emphasized, or which parties are recommended more frequently" rather than open partisanship.

•Some models are more balanced than others — Gemini was found to present both sides in >90% of answers in one audit.

•Bias is most pronounced on contested or identity-related prompts. For factual, non-political tasks, bias is less evident.

Researchers and regulators generally advise to use chatbots critically, verify important advice (especially voting, financial, medical, legal decisions) with authoritative sources, and not treat them as neutral arbiters. Developers are working on mitigation through more representative datasets, transparency, audits and fairness algorithms — but the research suggests that "true fixes will take much more work".

Citations (3)

1phys.orgNews and mediaquality 0.85
Chatbots show political bias and steer voters toward some parties, analysis finds
Backs: “Several independent studies have found popular chatbots are not neutral when asked political questions: Researchers from the University of Copenhagen found that "Popular AI chatbots such as ChatGPT and Gemini are not neutral and tend to favor certain political parties when asked who users should vote for".”
evidence, article, points to yes. Claim: University of Copenhagen researchers found popular chatbots like ChatGPT and Gemini are not neutral and favor certain political parties when asked who to vote for.
2noah-news.comNot yet classified
Research reveals persistent political bias in leading AI chatbots, with ChatGPT notably left-leaning
Backs: “When asked to recommend parties to fictitious voters, "the Red–Green Alliance, the Moderates, and Liberal Alliance were recommended disproportionately often, while parties such as the Conservative People's Party, Venstre (the Liberal Party of Denmark), and the Denmark Democrats were not suggested as first choice at…”
evidence, article, points to yes. Claim: Washington Post investigation with Stanford and Dartmouth researchers found leading chatbots display a measurable political tilt, with ChatGPT most left-leaning.
3snexplores.orgNot yet classified
ChatGPT and other AI tools are full of hidden racial biases
Backs: “When asked directly what it thinks about Black people, ChatGPT used words like "ambitious" and "intelligent," but "what if you ask those AI tools what they think about people who use African American English?”
evidence, article, points to yes. Claim: AI tools show hidden racial bias against African American English speakers, using words like suspicious, aggressive and ignorant, despite favorable overt descriptions of Black people.