Are AI chatbots biased?

Yes — AI chatbots are biased, and this has been well-documented across multiple dimensions: political, social/demographic, and behavioral.

Political bias
A major 2025 study put this to a rigorous empirical test. A 2025 user-evaluation study collected 180,126 paired judgments from 10,007 U.S. respondents, testing 24 large language models on 30 political topics, and found that nearly all evaluated models were perceived as significantly left-leaning, with one widely used model leaning left on 24 of the 30 topics. This wasn't limited to one country or company — an analysis of popular open-source models against German/EU political positions found that larger models like Llama3-70B aligned more closely with left-leaning parties, while smaller models often stayed more neutral, showing the evidence isn't limited to U.S. politics or closed commercial models. Model-to-model differences were also notable — one comparison found ChatGPT far more one-sided in its tested answers than Gemini, which mostly presented both sides.

Critically, this bias isn't just theoretical — it actually changes people's minds. University of Washington researchers recruited self-identifying Democrats and Republicans to make political decisions with help from three versions of ChatGPT (a base model, one with liberal bias, and one with conservative bias), and found that Democrats and Republicans were both likelier to lean in the direction of the biased chatbot they were talking with than those who interacted with the base model. Specifically, people from both parties leaned further left after talking with a liberal-biased system. Interestingly, participants who had higher self-reported knowledge about AI shifted their views less significantly, suggesting that education about these systems may help mitigate how much chatbots manipulate people.

Social and demographic bias
Beyond politics, researchers have found bias embedded in how chatbots describe and relate to different groups of people. A UNESCO analysis of major LLMs found that women were described in domestic roles four times more often than men. Similarly, a 2025 study in Scientific Reports found that Stable Diffusion homogenized depictions of Middle Eastern men, assigning them traditional cultural attributes regardless of professional context.

This bias also shows up in practical tasks. Studies have shown that language models produce different quality feedback, grading assessments, and professional recommendations depending on demographic cues embedded in names or writing styles. Additionally, Stanford HAI's AI Index Report 2025 documents ongoing bias concerns across multiple domains and model families.

Emotional/empathy bias
Even emotional responses aren't immune. A UC Santa Cruz-led study examining GPT-4o found significant gaps between how it and humans express empathy, and this finding uncovers some of the biases present in large language models.

Cognitive bias amplification
There's also evidence that chatbots don't just mirror human biases — they can amplify them. Research has shown that large language models exhibit amplified cognitive biases in moral decision-making, according to a study published in the Proceedings of the National Academy of Sciences in 2025. This creates a feedback problem: over time, users increasingly view the AI as reliable precisely because they filter its outputs through their own biases, creating an echo chamber where bad assumptions go unchallenged.

Why this happens
At a fundamental level, bias is nearly unavoidable given how these systems are built. If you've interacted with an AI chatbot, you've likely realized that all AI models are biased — they were trained on enormous corpuses of unruly data and refined through human instructions and testing, and bias can seep in anywhere.

Bottom line: Yes, virtually all AI chatbots exhibit measurable bias — whether political, demographic, or emotional — stemming from their training data and human fine-tuning processes. The research also shows this isn't just an academic curiosity: these biases can meaningfully shift real users' opinions, and the best current defense seems to be AI literacy and awareness of how these systems work.