Are AI chatbots biased?

Are AI chatbots biased? I'll sort the evidence on political, gender, and racial bias in today's models.

Yes, research indicates AI chatbots can be biased — but the nature, degree, and causes of that bias are nuanced.

1. What the research shows: Yes, but not uniformly

A careful answer to "Are AI chatbots biased?" is yes, but the type of bias depends on the model, the task, and how bias is measured.

a) Political bias - consistently left-leaning in tests

Multiple peer-reviewed and journalistic tests have found leading large language models (LLMs) tend to produce left-of-center responses on politically charged questions.

Importantly, bias is not always perceived by users. One study found most respondents viewed chatbots as politically neutral rather than showing partisan bias (62% vs. 38%), but among those who did perceive bias, views were split.

b) Social biases - gender, race, and other human stereotypes

Chatbots reflect biases present in human language:

A recent example is salary advice:

c) Other documented biases

This unintended power to persuade is caused by "latent biases introduced during the training of the large language models (LLMs) that drive chatbots' core capabilities" which "can carry over from ideological leanings in the data used to train LLMs" and "lend subtle nuances to the framing of the narratives the chatbots generate". Compared to Wikipedia entries, both default AI summaries and those prompted with liberal framing "caused participants to express more liberal opinions", though effects were modest - representing a slight difference from leaning moderate to somewhat liberal.

2. Why are chatbots biased?

There is no evidence that most mainstream developers intentionally code left-wing or right-wing ideology into models. Rozado acknowledged that chatbots' beliefs were likely not coded deliberately.

Instead, bias emerges from:

1. Biased training data: "Biased Training Data: Chatbots learn from the data they are provided, and biases in that data will be transferred to the chatbot's behavior". For example, if a chatbot is trained on text data that is biased against women, the chatbot will learn that behavior.
More generally, "AI chatbots are only as unbiased as the data they are trained on. However, training datasets often reflect societal biases, and chatbots can unintentionally replicate and amplify these biases", because datasets are often sourced from the internet.

There are two ways bias can creep in: the people collecting data collect a biased sample not representative of the wider data set, or the wider data set itself is biased.

2. Model training and alignment: After pre-training, models go through supervised fine-tuning and reinforcement learning from human feedback (RLHF). The preferences of human raters, as well as safety guidelines designed to avoid harmful or offensive outputs, can shift the model's average response style. This helps explain why base models without fine-tuning showed less consistent political lean, but gave answers with "little link to the questions being asked", while fine-tuned conversational models showed clearer left-leaning tendencies.

3. Memory and personalization: Bias does not require you to state your identity. Researchers warned that LLMs remember conversations over time, so if you mention you are a woman or a refugee, that context can inform future advice, making personalization become "invisible discrimination when you ask for salary negotiating tactics".

3. How big is the problem?

4. What is being done and what can users do?

Developers including OpenAI, Google, and Anthropic have acknowledged bias issues and conduct internal audits, adjust training data curation, and refine RLHF guidelines. Google, for example, temporarily halted Gemini's image generation to correct racial and gender bias concerns.

For users:

In short, AI chatbots are not neutral mirrors of reality. They inherit and sometimes amplify patterns — political, gender, racial, and cognitive — present in the human-generated data they were trained on, plus the values embedded during human feedback training.