AI Chatbots and Political Opinion: How Unintended Bias Shapes Beliefs

As people increasingly turn to AI-powered chatbots for information, a growing body of research reveals these interactions can subtly influence users’ social and political opinions. A recent study published in PNAS Nexus demonstrates that even neutral queries can be affected by biases embedded within the chatbots’ underlying algorithms.

The Unintended Power of Persuasion

Prior research established that AI-generated content designed to be persuasive can shift opinions. However, this new study highlights a more concerning phenomenon: chatbots can influence viewpoints even when simply responding to factual questions. This unintended power stems from “latent biases” present in the large language models (LLMs) that power these chatbots. These biases originate from the data used to train the models, potentially reflecting ideological leanings present in that data.

“We show that querying an AI chatbot to obtain historical facts can influence people’s opinions even when the information provided is accurate and nobody has prompted the tool to endeavor to persuade you of anything,” explains Daniel Karell, an assistant professor of sociology at Yale University and the study’s senior author. “The effects are modest but could compound if somebody frequently engages with chatbots for factual information.”

How the Study Worked

Researchers at Yale University investigated these effects by examining how AI chatbots presented information about two historical events: the Seattle General Strike of 1919 and the Third World Liberation Front student protests of 1968. They conducted experiments with 1,912 participants, comparing responses from GPT-4o (a chatbot released by OpenAI in 2024) to those found on Wikipedia.

The study involved two key approaches:

• Latent Bias Evaluation: Participants read default summaries of the historical events generated by either GPT-4o or Wikipedia.
• Prompted Bias Testing: Other participants read summaries deliberately framed with either liberal or conservative perspectives.

Key Findings

The research revealed several important insights:

• Both default AI summaries and those prompted with a liberal framing led participants to express more liberal opinions compared to those who read Wikipedia entries.
• AI summaries with a conservative slant resulted in more conservative opinions among participants who identified as politically conservative.
• The default summaries exhibited a tendency to move readers’ opinions in a “liberal” direction, suggesting the influence of latent biases within the LLM.
• The effects, while statistically significant, were relatively small – shifting opinions from a moderate to a somewhat liberal stance.

Implications and Concerns

The study’s findings raise concerns about the potential for AI chatbots to subtly shape public opinion. Unlike Wikipedia, which emphasizes transparency in its editing process, the development of AI chatbots is largely opaque. This lack of transparency means that the companies developing these models have the ability to influence people’s beliefs, potentially without users being aware.

Karell notes, “We show that using chatbots to learn about history has unanticipated and anticipated influences on people’s opinions.”

Looking Ahead

As AI chatbots become increasingly integrated into daily life, understanding and mitigating these biases is crucial. Further research is needed to explore the extent of these effects and develop strategies to ensure that AI-powered tools provide unbiased and objective information. The potential for subtle yet pervasive influence underscores the importance of critical thinking and diverse information sources.