An online conversation about US gun control might seem an unlikely place to test whether AI can improve political debate. Yet a field experiment with 1,574 participants found that offering people more considerate wording could make discussions less divisive while leaving their policy views unchanged.
A wording suggestion before each message
Participants with different views on US gun control discussed the issue in an online chat room. Before a message was sent, GPT-3 could suggest an alternative phrasing intended to sound more reassuring, validating, or polite.
Participants remained in control of what they said. They could adopt the suggestion, edit it, or send their original message. Across the experiment, GPT-3 made 2,742 suggestions, and participants accepted 1,798 of them—about two-thirds. Conversations averaged twelve messages, and participants were paid.
Better discussion, without changing policy views
The researchers assessed the conversations with text analysis, comparing messages before and after an AI suggestion. They also used questionnaires completed before and after the chats. The results supported their hypothesis: AI phrasing assistance improved the quality of conversation and reduced divisiveness.
Chris Rytting, a participating researcher and a Ph.D. candidate at Brigham Young University, said the suggestions had “significant effects in terms of decreasing divisiveness and increasing conversation quality, while leaving policy views unchanged.” The team reported that participants’ political opinions were not affected.
That distinction matters. The experiment tested how people expressed disagreement, not whether they would change their positions on gun control. A discussion can become more respectful while its participants continue to disagree.
Why the tone of online debate matters
The team said the study was prompted by the rapid escalation of debates on digital channels and the question of whether AI could help improve discussion culture. The paper warns that toxicity can increase polarization and weaken the ability of diverse societies to work together on social problems.
For the researchers, civil dialogue has value even when it does not settle an argument. They describe it as a necessary, though insufficient, condition for greater mutual understanding, compromise, and coalition building. In practical terms, a calmer exchange may give people more room to hear one another, but better wording alone cannot guarantee agreement or cooperation.
What the experiment leaves open
A follow-up survey conducted nearly three months later checked whether the chatbot’s effects lasted. The researchers found no anomalies in that follow-up, but the study’s authors said more research is needed to confirm and extend the findings in real-life conditions.
They pointed to longer, unpaid conversations among family or friends as possible settings for future studies. Those situations would differ from the paid online chats, where discussions averaged twelve messages. Further research could show whether phrasing suggestions help in conversations that unfold over more time and within existing relationships.
For now, the experiment offers evidence that a language model can influence the tone of a political discussion without shifting policy views. It also leaves a larger question for future work: whether that kind of assistance can make a lasting difference outside a short, structured chat.