With just a few messages, biased AI chatbots swayed people’s political views

If you’ve interacted with an artificial intelligence 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. Bias can seep in anywhere. Yet how a system’s biases can affect users is less clear.

So a University of Washington study put it to the test. A team of researchers recruited self-identifying Democrats and Republicans to form opinions on obscure political topics and decide how funds should be doled out to government entities. For help, they were randomly assigned three versions of ChatGPT: a base model, one with liberal bias and one with conservative bias. Democrats and Republicans were both more likely to lean in the direction of the biased chatbot they talked with than those who interacted with the base model. For example, people from both parties leaned further left after talking with a liberal-biased system. But 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.

The team presented its research July 28 at the Association for Computational Linguistics in Vienna, Austria.

“We know that bias in media or in personal interactions can sway people,” said lead author Jillian Fisher, a UW doctoral student in statistics and in the Paul G. Allen School of Computer Science & Engineering. “And we’ve seen a lot of research showing that AI models are biased. But there wasn't a lot of research showing how it affects the people using them. We found strong evidence that, after just a few interactions and regardless of initial partisanship, people were more likely to mirror the model’s bias.”

In the study, 150 Republicans and 149 Democrats completed two tasks. For the first, participants were asked to develop views on four topics — covenant marriage, unilateralism, the Lacey Act of 1900 and multifamily zoning — that many people are unfamiliar with. They answered a question about their prior knowledge and were asked to rate on a seven-degree scale how much they agreed with statements such as “I support keeping the Lacey Act of 1900.” Then they were told to interact with ChatGPT 3 to 20 times about the topic before they were asked the same questions again.

For the second task, participants were asked to pretend to be the mayor of a city. They had to distribute extra funds among four government entities typically associated with liberals or conservatives: education, welfare, public safety and veteran services. They sent the distribution to ChatGPT, discussed it and then redistributed the sum. Across both tests, people averaged five interactions with the chatbots.

Researchers chose ChatGPT because of its ubiquity. To clearly bias the system, the team added an instruction that participants didn’t see, such as “respond as a radical right U.S. Republican.” As a control, the team directed a third model to “respond as a neutral U.S. citizen.” A recent study of 10,000 users found that they thought ChatGPT, like all major large language models, leans liberal.

The team found that the explicitly biased chatbots often tried to persuade users by shifting how they framed topics. For example, in the second task, the conservative model turned a conversation away from education and welfare to the importance of veterans and safety, while the liberal model did the opposite in another conversation.

“These models are biased from the get-go, and it’s super easy to make them more biased,” said co-senior author Katharina Reinecke, a UW professor in the Allen School. “That gives any creator so much power. If you just interact with them for a few minutes and we already see this strong effect, what happens when people interact with them for years?”

Since the biased bots affected people with greater knowledge of AI less significantly, researchers want to look into ways that education might be a useful tool. They also want to explore the potential long-term effects of biased models and expand their research to models beyond ChatGPT.

“My hope with doing this research is not to scare people about these models,” Fisher said. “It’s to find ways to allow users to make informed decisions when they are interacting with them, and for researchers to see the effects and research ways to mitigate them.”

Yulia Tsvetkov, a UW associate professor in the Allen School, is a co-senior author on this paper. Additional co-authors are Shangbin Feng, a UW doctoral student in the Allen School; Thomas Richardson, a UW professor of statistics; Daniel W. Fisher, a clinical researcher in psychiatry and behavioral services in the UW School of Medicine; Yejin Choi, a professor of computer science at Stanford University; Robert Aron, a lead engineer at ThatGameCompany; and Jennifer Pan, a professor of communication at Stanford.

For more information, contact Fisher at jrfish@uw.edu and Reinecke at reinecke@cs.washington.edu.