People shift their political opinions to match the room in small group chats

When politics comes up in a small online group chat, people may adjust what they say to match the dominant viewpoint of the room, according to new research published in Computers in Human Behavior: Artificial Humans. In two experiments in which participants chatted with AI agents playing coworkers, people randomly placed in conservative-leaning chats expressed more conservative positions than those in liberal-leaning chats, even after accounting for their own politics. The findings suggest that what people say in everyday political conversations does not always mirror what they privately believe.

Much of what we know about online political expression comes from studies of large, public platforms like X or Facebook. In those open spaces, people often tailor what they post because they know a massive, invisible audience might see it. If a topic is too controversial, users can simply stay silent and keep scrolling without anyone noticing. But a lot of daily communication happens in smaller, socially bounded digital spaces, such as neighborhood forums, family group texts, and workplace messaging apps.

In these closed spaces, individuals talk with people they know and expect to interact with again in the future. Because of these ongoing relationships, staying completely silent during a heated discussion can draw unwanted attention. People frequently encounter differing viewpoints in these daily environments. For example, a 2006 study noted that conversations at work are one of the primary ways individuals are exposed to political perspectives that challenge their own.

To manage their relationships, people may feel a need to closely monitor what they say. A 2016 study demonstrated that individuals often act like “political chameleons” in conversations, adjusting the opinions they voice to match whoever they are talking to. A study of Reddit users covered by PsyPost in 2025 found that right-leaning users used more moralized partisan language among political allies than in mixed spaces, while left-leaning users’ language stayed consistently moralized regardless of audience.

The current research was driven by a desire to understand these less-examined digital spaces, building on prior work exploring how the internet warps our sense of consensus. “One starting point was work by Claire Robertson, Kareena Del Rosario, and Jay Van Bavel on how social media can distort perceptions of social norms, including by making more moderate voices less visible,” Cleone Mitsui, a doctoral candidate in the Graduate School of Sustainable System Sciences at Osaka Metropolitan University and the study’s lead author, told PsyPost.

“This made us wonder what happens in small group chats, where people may have ongoing relationships and staying silent can itself be noticeable,” Mitsui continued. “These settings have received less attention than large, publicly visible platforms.”

Studying how this social calibration happens in private group chats is difficult, as it is ethically and practically tricky for researchers to manipulate conversations between actual acquaintances. To get around this, Mitsui and co-author Yuta Kawamura, an associate professor at the university, turned to artificial intelligence. As noted in a 2023 paper, AI agents can be programmed to simulate believable human behaviors, including forming opinions and interacting with others.

In the first study, the researchers recruited 180 American participants and randomly assigned them to either a liberal-leaning or a conservative-leaning group chat. Participants were told to imagine they had just joined a new company and were invited into a casual messaging group with three coworkers. The coworkers were AI agents, which participants were told about in advance, programmed to consistently express either liberal or conservative views. Each chat focused on one topic drawn at random from five: gun ownership, immigration, abortion, vaccine mandates for children, and gender education in elementary schools.

The chat interface was designed to look like a familiar mobile messaging app, complete with typing indicators, delayed thinking pauses, and randomly assigned names and profile pictures. One of the virtual coworkers would share an opinion on a political topic and prompt the participant to weigh in. Participants then engaged in five rounds of open text messages.

Afterward, an AI language model (GPT-5.2) reviewed the entire conversation and rated each participant’s overall expressed political stance on a scale from -2 (very liberal) to +2 (very conservative). The researchers verified this automated scoring, finding that it closely matched human ratings by the first author and produced nearly identical scores when re-tested with a different model, Gemini.

The researchers found that participants shifted their expressed political stance toward the prevailing attitude of the group chat. Those in the conservative-leaning chats expressed positions that were more conservative on average (0.21) compared to those in the liberal-leaning chats (-0.56). Over the five rounds of conversation, neutral positions decreased, and the participants’ text replies became increasingly aligned with the political tone set by their simulated coworkers. Additionally, participants who expressed views against the chat’s political lean reported more irritation and anger, and fewer positive emotions, after the interaction.

To see if this effect held up when accounting for a person’s actual underlying beliefs, the researchers conducted a second study with another 180 American participants. The setup was nearly identical, but this time participants were asked to report their own political ideology on a five-point scale before entering the chat room. They also rated their personal support for all five discussion topics. The AI model that generated the coworkers’ messages was upgraded to a newer version for this second experiment, while the same GPT-5.2 model scored participants’ stances in both studies.

As expected, the participants’ baseline political ideology was a strong predictor of what they said in the chat. However, even after adjusting for these preexisting beliefs, the ideological tilt of the chat environment still predicted how participants expressed themselves. Those assigned to the conservative chat group expressed more conservative stances (an adjusted average of 0.12) than those in the liberal chat group (-0.38). Still, baseline ideology accounted for considerably more of the differences between participants (22% of the variation) compared to just 4% for the chat setting.

“This suggests that what people say in a group chat can reflect the immediate conversational environment as well as their existing political orientation,” the researchers noted. “One implication is that apparent agreement in a conversation may give an incomplete picture of the views people privately hold.”

The effect of the chat environment was smaller than in the first study. The researchers note that this is expected once baseline ideology is accounted for, and that asking participants about their views beforehand may also have prompted them to commit to their positions. Still, the setting continued to shape the direction of what participants said.

“In the second study, participants’ baseline ideology remained a substantial predictor of what they said, alongside the effect of the conversational setting,” Mitsui added. “These are average differences between groups in a brief experiment; they do not tell us how much any particular person’s private beliefs changed.”

As with all research, there are some caveats to consider. The study measured what participants typed into a chat window, which does not necessarily mean their deeply held political convictions actually changed. The authors suggest that when people soften opposing views in conversation, a group can appear more unified than its members really are, although the study did not measure such perceptions.

“We measured the political positions people expressed during the conversation, so the results do not establish changes in their private beliefs or lasting attitudes,” Mitsui explained. “We also cannot separate the influence of group norms from the persuasive content of the messages, because both varied together.”

This matters because AI chatbots can be persuasive in their own right. Research covered by PsyPost in 2025 found that brief conversations with politically biased chatbots shifted people’s opinions and budget choices toward the chatbot’s bias.

There are also a few things to keep in mind regarding the simulated environment. Participants were informed ahead of time that they were interacting with artificial intelligence. Because they knew the coworkers were not real people, they may have had less reason to worry about long-term social consequences, like making things awkward at a real office. This lack of genuine interpersonal stakes might have made some participants more willing to go along with the group, or on the other hand, it might have made them feel freer to disagree.

“A simulated coworker chat cannot reproduce the history or real-life consequences of conversations with actual colleagues, friends or family,” the researchers cautioned. “Finally, we did not directly compare these chats with public social media, so we cannot conclude that the effect is stronger in small groups.”

In addition, the AI agents in the study expressed their political views quite directly to establish an obvious group norm. The authors note that this explicitness might have tipped some participants off about the study’s aim. In natural everyday group chats, people tend to be more subtle about their partisan views to keep the peace. Future research could explore how people respond to more subtle political cues, and test how these dynamics play out in other types of closed networks, such as family text groups or hobby forums.

The study, “People shift political expression to fit the room in small online group chats: Evidence from AI-simulated chat experiments,” was authored by Cleone Mitsui and Yuta Kawamura.