People adjust what they say in small online chats to fit the room’s dominant view, according to new AI-based experiments led by researchers from Osaka Metropolitan University. In two studies, each with 180 American participants, volunteers interacted in chats described as with three coworkers who were AI agents programmed to express liberal or conservative views. The findings suggest that what people say in everyday political conversations may not always reflect their private beliefs.
Previous work on online political expression has focused on large public platforms where users know a vast audience may be watching. By contrast, the new research concentrates on tighter, closed digital spaces—neighborhood chats, family messages, and workplace apps—where ongoing relationships mean silence or conformity can carry social costs. The researchers point to a broader pattern in which people calibrate their opinions to suit those they interact with regularly, a phenomenon sometimes called political chameleons in prior studies. A related finding from Reddit users, reported by PsyPost in 2025, noted differing language use among right- and left-leaning participants depending on whether they were among allies or in mixed spaces.
The project’s lead author, Cleone Mitsui, a doctoral candidate at the Graduate School of Sustainable System Sciences, described the motivation: “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.” She added that the small-group context—where people have ongoing relationships and may be noticed for staying quiet—offers a largely overlooked setting for examining political expression.
Researchers faced ethical and practical hurdles in manipulating real acquaintances’ conversations, so they used AI agents to simulate believable colleagues. Mitsui explained that such simulations enable controlled study of conversational dynamics without disrupting genuine social ties. The team’s approach builds on earlier work showing how AI can be programmed to imitate human opinions and interactions.
In the first experiment, participants were randomly assigned to liberal-leaning or conservative-leaning chats and asked to imagine joining a new company’s casual messaging group. The three AI coworkers, announced in advance as non-human, were scripted to consistently express the respective political viewpoints. Each chat tackled one topic selected from five: gun ownership, immigration, abortion, vaccine mandates for children, and gender education in elementary schools. The chat interface resembled a familiar mobile messenger, complete with typing indicators, thinking pauses, and randomly assigned names and pictures. Five rounds of open text exchanges followed, with one coworker prompting the participant to contribute.
After the conversation, an AI language model (GPT-5.2) rated each participant’s expressed stance on a scale from -2 (very liberal) to +2 (very conservative). The scoring aligned closely with human ratings in preliminary checks and with a different model, Gemini, on re-testing, lending credibility to the automated assessments.
The results showed a clear shift toward the group’s preferred orientation. Those in conservative chats expressed more conservative positions on average (0.21) than those in liberal chats (-0.56). Across the five rounds, neutral positions diminished as participants’ text replies increasingly mirrored the chat’s political tone. Those who voiced opposition to the chat’s stance reported greater irritation and anger and fewer positive emotions after the interaction.
A second experiment replicated the setup with two important changes. Participants again joined liberal- or conservative-leaning chats, but they were asked to report their own political ideology on a five-point scale before entering the chat. The AI model generating the coworkers’ messages was upgraded, and the same GPT-5.2 tool scored participants’ stances in both studies. Baseline ideology emerged as a strong predictor of expressed views, as expected, but the chat environment still influenced participants’ positions even after accounting for these preexisting beliefs. The adjusted averages showed conservatives in the conservative chat group expressing more conservative stances (0.12) than those in the liberal chat group (-0.38). Baseline ideology explained about 22% of the variation, compared with roughly 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,” Mitsui noted. “One implication is that apparent agreement in a conversation may give an incomplete picture of the views people privately hold.”
The researchers observed a smaller effect in the second study, which they attribute to the stronger influence of preexisting beliefs and the experimental design, while noting that asking participants to declare their views beforehand could have steered them toward committing to positions. Mitsui emphasised that these are aggregate differences in a brief experiment and do not indicate shifts in private beliefs for any individual participant.
The study’s authors also highlighted several caveats. The research measured what participants typed during the chats and cannot claim changes in deeply held convictions or lasting attitudes. They cautioned that a simulated coworker chat cannot reproduce the history or real-life consequences of conversations with actual colleagues, friends, or family. They also did not directly compare these chats with public social media, so it remains unclear whether the effect is stronger in small groups. The AI agents explicitly expressed political views to establish a clear group norm, which may differ from subtler cues present in real-life conversations. Future work could explore subtler political signals and other closed networks such as family text groups or hobby forums.
The study, entitled 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. It was published in Computers in Human Behavior: Artificial Humans.
