Are AI chatbots biased?

Yes — research consistently shows AI chatbots exhibit various forms of bias. Here's what the evidence reveals across multiple dimensions:

Social and Demographic Bias

Studies have found chatbots reproduce human-like prejudices. One study found that AI chatbots did produce biased content, and it wasn't even borderline, happening in the AI equivalent of the subconscious in response to innocent prompts. In a comparison with human-written news, researchers found the AI language models were much worse than imperfect news outlets, ranging in some cases from 40% to 60% more biased against minorities in their language choice. The same analysis found the sentiment of the language was consistently more toxic.

A separate study mirroring classic human bias experiments concluded that large language models such as GPT-3 often reflect human biases and are apt to use gender stereotyping and to focus on threat, negativity and gossip. Across five different experimental designs, GPT-3 produced broadly the same biases as humans, leading researchers to conclude "AI is not a neutral agent," which matters because many organizations are considering using AI to write press releases and summarize scientific articles.

Research also shows chatbots can be less helpful to certain users. MIT researchers found that these AI systems may actually perform worse for the very users who could most benefit from them, highlighting growing risks especially for marginalized communities worldwide.

Political Bias

This is one of the most studied and debated areas. A 2025 Washington Post analysis found that AI models behind ChatGPT, Gemini and others, when tested with political questions designed by researchers, show results suggesting chatbots have clear political leanings that can conflict with promises made by the companies behind them. The Post noted several previous academic studies found that AI models powering chatbots tend to favor left-leaning positions.

A large-scale 2025 study was particularly direct: it 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. A related Stanford study found that both Republicans and Democrats perceived left-leaning bias in LLMs' discussion of contentious topics — though importantly, prompting a model to adopt a neutral stance generated responses that users found less biased and considered higher quality, meaning bias can often be mitigated.

European research found similar patterns: one study examining ChatGPT and Gemini in EU election contexts found ChatGPT exhibited a strong tendency to give positive ratings to left-wing and centrist political groups, while right-wing and populist parties consistently received low ratings, along with a clear preference for pro-European attitudes.

Why This Matters: Persuasion Effects

Perhaps most concerning, biased chatbots don't just reflect opinions — they can actively shift them. A University of Washington study found that biased AI chatbots swayed people's political views with just a few messages, and people from both parties leaned further left after talking with a liberal-biased system. Notably, participants with higher self-reported AI knowledge shifted their views less significantly, suggesting education about these systems may help mitigate manipulation.

A large 2025 Nature study reinforced this: Trump supporters who talked with a pro-Harris bot shifted their views in her favor by almost four points, and Harris supporters who chatted with a pro-Trump bot altered their views in his favor by more than two points. When repeated internationally, the effects were even larger in Canada and Poland, with chatbots shifting voters' candidate ratings by ten points on average.

Root Causes

Researchers trace these biases largely to training data and human feedback processes: AI models' biases largely stem from the data on which they are trained, and can be amplified during both training and inference.

Bottom Line

The consensus across academic literature is that AI chatbots are not neutral — they inherit biases from training data, internet text, and human feedback processes (RLHF), manifesting as social/demographic bias, political slant, and disparities in how they serve different user groups. However, this isn't necessarily permanent or unfixable: studies show bias can be measured, and techniques like neutrality prompting or targeted bias-mitigation strategies can reduce it, as demonstrated by projects achieving high-quality, unbiased academic advising support with no stereotypical biases found in the sample data through careful prompt engineering.