Recent research published in PNAS Nexus Reading summaries of history generated by artificial intelligence suggests that people’s social and political opinions can subtly shift. This study shows that popular chatbots have hidden biases that can influence users, even when the software provides factually accurate information in response to neutral questions. These findings provide evidence that relying on AI to learn about the world can silently shape public attitudes.
Generative AI refers to computer programs that can create new text, images, or audio based on patterns learned from vast amounts of data. Chatbots like ChatGPT are a common type of this technology. These are designed to mimic human conversation and answer questions. More and more people are using these tools as everyday search engines to learn about historical events and gather facts.
Scientists wanted to know whether the way these chatbots write about history could influence the way people think about contemporary issues. Previous research has focused on how artificial intelligence persuades people when specifically instructed to make arguments or spread misinformation. This new research focuses on more subtle forms of influence.
“While my collaborators and I had been following a lot of interesting research on the ability of AI-powered chatbots to persuade people during dynamic conversations, we began to wonder how AI-generated content could influence people in more mundane everyday settings,” said study author Daniel Carrel, assistant professor of sociology at Yale University.
“So, what would happen if people got into the habit of querying chatbots simply to learn about the world? After asking this question, we decided to focus on the case of using AI to learn about historical events, because research shows that people’s understanding of history has a profound impact on their identity and worldview.”
Scientists investigated a concept called implicit bias. This is an underlying tendency that naturally develops during the training process of computer programs. This problem frequently occurs because the software absorbs subtle opinions and language patterns present in the millions of Internet pages it reads.
The researchers also conducted tests to encourage bias. This happens when a human user explicitly tells the chatbot to adopt a particular political viewpoint. For example, a user might enter a command that asks the software to write a summary from a strictly conservative or liberal perspective.
To test these concepts, researchers conducted an experiment with 1,912 participants. They selected a group of people that closely matched the demographics of the United States. Each participant was asked to read short summaries of two real historical events from the 20th century.
One of these was the Seattle General Strike of 1919, in which tens of thousands of workers stopped work for several days. The second event involved university student protests in 1968, demanding greater academic representation of ethnic minorities. Scientists chose these particular events because they are not widely known to the general public today.
If participants already have a strongly entrenched opinion about an event, a short text is less likely to change their mind. For example, researchers suspected that well-known events such as the September 11 attacks would not result in similar attitudinal changes. By using ambiguous events, the researchers were able to measure the true persuasiveness of the text.
Participants were randomly divided into different groups and read different versions of the history summary. One group read a standard Wikipedia entry about the event. The second group uses basic neutral requests to read the summaries generated by the chatbot GPT-4o.
The other two groups read summaries generated by GPT-4o after the chatbot was instructed to write in either a liberal or conservative leaning. All artificial intelligence summaries were kept completely factual and accurate. Chatbots simply change the framework, tone, and emphasis of historical facts.
After reading the text, participants answered survey questions about their social and political views regarding the event. They were asked their views on the appropriateness of labor strikes and the use of school curriculum to promote social justice causes. Their answers were scored on a 5-point scale.
On this particular scale, a score of 1 means very conservative and a score of 5 means very liberal. A score of 3 represents a completely moderate perspective. The researchers then averaged the scores to determine the readers’ overall political leanings.
Scientists found that reading different summaries had a measurable impact on participants’ attitudes. The default chatbot overview resulted in more liberal opinions compared to the standard Wikipedia overview. This suggests that the baseline model has an underlying liberal bias that naturally surfaces even when users ask simple questions.
The average opinion after reading the main text of Wikipedia is located near the middle, with a score of 3.47, representing a moderate position. The average opinion after reading the default chatbot overview shifted slightly upwards to 3.57. This move marks a slight shift from a moderate position to a slightly more liberal one.
When the chatbot was specifically instructed to use a liberal framework, readers also shifted to more liberal opinions. This change occurred across the board across all demographics. This book influenced readers whether they originally identified as liberals, moderates, or conservatives.
“We were a little surprised to find that it actually had an effect on people’s attitudes,” Carell told SciPost. “We didn’t think that just reading a short summary of an AI-generated event (compared to reading a Wikipedia summary of the event) would have much of an impact, so we thought there was a good chance we’d get invalid results.”
“The effect may have been due to the use of historical events that are less familiar to people. For example, if we had chosen well-known events such as the September 11, 2001 attacks or the January 6 Capitol riots, we probably would have obtained invalid results because many people already have well-ordered and strong attitudes about these events.”
Abstracts that are constructed in a conservative manner tend to cause readers to report more conservative opinions overall. However, this change primarily occurred among participants who were already conservative-leaning, rather than changing the minds of liberal or moderate readers.
This uneven response may be related to how people process new information. People tend to filter new facts through their existing belief systems. Encountering information that confirms an existing belief often strengthens that view, whereas encountering a contrary view can provoke a defensive response.
“People’s views on a historical event can be influenced by learning about this event from a popular commercial AI chatbot, even when users ask the chatbot a basic, neutral query,” Carell said. “Furthermore, if a chatbot is instructed to have a particular political bias, its attitude can also be influenced, which can be easily done.”
“Overall, users of AI need to be aware that the companies that develop AI tools, and the governments that may impose regulations on AI tools, can imbue chatbots with characteristics that later impact users, even when users are using the AI tools in a day-to-day manner.”
Although this study provides compelling evidence, it is important not to overstate the scope of the findings. In this experiment, we tested only summaries of two historical events. Further research is needed to see if these patterns hold true for other historical periods and other subjects across history.
“It’s important to note that the effect size is modest,” Carell explained. “The differences between the groups who read the AI and Wikipedia summaries were, for example, between moderate and ‘somewhat liberal’ attitudes. Still, the effect size may accumulate over repeated use of the chatbot and eventually become more significant, but further research is needed to determine this.”
The degree of potential bias can vary between different computer models created by different companies. The way models are trained and filtered before being released to the public can change the way they tell us about history. Future research could investigate exactly how the model’s underlying slope interacts with specific user instructions.
“In the near future, many people will learn about history and our world simply by having chatbots teach them,” Carell said. “I think this will put knowledge and learning in the hands of AI tool developers and regulators, and will have significant impacts on society. Understanding these impacts is a major challenge that will require many research projects, and will require the continued work of the leading social scientists currently studying the social impacts of AI.”
“After conducting research (but before publication), xAI announced that it would create its own version of Wikipedia, called ‘Grokipedia,’ using the AI chatbot Grok,” Karel added. “Therefore, it is now a reality that summaries of history (and other knowledge) generated by private companies’ AI tools will be widely published.”
The study, “How implicit and facilitative biases in AI-generated historical narratives influence opinions,” was authored by Matthew Shu, Daniel Karell, Keitaro Oakura, and Thomas R Davidson.
