early findings with experimental participants
We're using Sway to study how AI can help people have better discussions across deep political and moral divides.
In an early pilot study, 120 participants spent 30 minutes on Sway discussing a polarizing topic such as the 2020 election, January 6, Israel-Palestine, or U.S. abortion policy.
We are planning a large randomized controlled trial where we will compare results from discussions like these with unmediated discussions. We are also running a large field study involving students in 2025 at over 30 colleges and universities.
Pilot Study Highlights
At the beginning of every chat, the AI discussion facilitator, Guide, introduced the discussion topic and announced the participants' positions.
We have been consistently amazed to see participants emerge from discussions of these extremely polarizing topics with feedback like this:
Guide introduces a discussion on the topic of abortion policy.
Topic: the 2020 national U.S. election
"This is probably the only way I would ever have a conversation like this. I really appreciated that I could ask Guide for fact-checks instead of saying ‘You are wrong’. The fact that it was mediated helped the tone to stay civil and moved us along."
Topic: the January 6 Capitol Riot
“I definitely think Guide was very useful in leading the discussion. When we got off track, Guide made sure to point us in the right direction. I think this study will be very useful”
Topic: Israel-Palestine
"In a way, I felt safe with guide. I'm sensitive, so having that slight security with the guide, I felt better to speak my mind a little more."
Topic: U.S. abortion policy
"I thoroughly enjoyed it. The guide was a great mediator that really help set the pacing and also control the discussion. Without the guide, I can see the conversation get messy with people talking over one another. It created a thought provoking conversation instead of two people rambling."
Topic: U.S. funding for Ukraine
"I feel like this is the best AI-driven anything that I've ever personally experienced. It asked very thoughtful, relevant, specific questions to direct the conversation. It was patient, encouraging, and respectful."
Topic: U.S. abortion policy
"It was actually amazing. Guide interjected at the right times without being overly intrusive. I went in thinking that I couldn’t compromise, but I left with a more open-minded view. I really enjoyed it."
Topic: Israel-Palestine
"I found Guide so useful in developing the discussion further. I think it made both parties delve deeper into some of the statements both made. I really enjoyed discussing this topic with someone else. I love the whole concept."
Topic: the January 6 Capitol Riot
"Not only did Sway guide our conversation in a thought provoking way, it moved us in a very civil direction. I was able to see things from my partner’s point of view. I also changed my stance!”
Both before and after their AI-guided discussions, participants rated how strongly they agreed or disagreed with a series of statements designed to measure their openness to opposing viewpoints.
On every statement, the overall effect was a statistically significant shift in the direction of more openness.
Here we see ratings of the statement, I feel like I can understand people who disagree with me about this topic.
After a 30-minute Sway discussion, nearly half of the participants (47%) gave higher ratings to the statement—and only 14% felt like they had less understanding of their opponents than they did before chatting.
Overall, participants' level of agreement increased by 0.41 standard deviations (95% confidence interval: 0.24 to 0.59). This difference was highly statistically significant: t(120) = 4.8, p < 0.00001.
Next we see participants' ratings of the statement, People who disagree with me about this topic have well-thought out reasons for their views.
61% of participants agreed more strongly with the statement after their Sway chats—that's more than twice as many who moved in the opposite direction (28%).
The improvement in participants' ratings before and after discussion was highly statistically significant: t(120) = 4.5, p = 0.0003.
Looking at the nine-item scale as a whole (Cronbach's α = 0.89) , we see that two-thirds of participants had higher scores after chatting with an opponent on Sway, while only one-in-five had lower scores.
This change was highly statistically significant: t(120) = 6.1, p < 0.00001.
And at 0.32 standard deviations (95% confidence interval: 0.21 to 0.44) it's also meaningfully large.
We expect these effects to deepen and become more ingrained after repeated discussions on Sway—a hypothesis we will test in our RCT.