Sway is a teaching tool, but it is also a research instrument: a round-the-clock laboratory in which large numbers of student volunteers contribute de-identified data from real conversations about scholarly as well as political, social, and moral topics. We use it to ask three connected questions.
Sway is also a research instrument: student volunteers contribute de-identified data from real conversations about scholarly, political, social, and moral topics. We use it to ask three questions.
Can guided practice with constructive disagreement — rather than instruction about it — cultivate intellectual virtues like humility and charity?Can guided practice with disagreement — not just instruction about it — cultivate humility and charity?
What makes facilitation work best? Given what we know about a topic and about the students discussing it, what should a facilitator do to promote an excellent discussion? Sway lets us study this empirically, in real conversations at every kind of post-secondary institution.What makes facilitation work best? Sway lets us study this empirically, in real conversations.
Do the skills transfer? ICDC, a FIPSE-funded national initiative led by researchers at North Carolina State and Notre Dame with instructors from over 100 campuses, is studying whether students who practice these skills on Sway internalize them and generalize them to in-person classroom discussions and to their lives beyond the classroom.Do the skills transfer? ICDC, a FIPSE-funded initiative at 100+ campuses, is testing whether skills practiced on Sway carry into classroom discussions and beyond.
We are preparing large randomized controlled trials that will compare AI-facilitated discussions with unmediated ones, and the FIPSE study will reach nearly 100,000 students at over 125 campuses over the next three years — making it the largest empirical study of civil discourse pedagogy to date. For conversation-level evidence today, see our student chat threads and instructor reports.
Most constructive dialogue programs can only report what participants say about themselves. Sway provides behavioral evidence. As one example, each student’s opinion is measured before and after each conversation. Students at the extremes moderate. Students with no opinion start forming their views.
13,992 conversations55% of extreme starters moderated4.3× depolarization‑to‑polarization ratio
A diverging bar chart. One row for each of the seven starting positions on the opinion scale, most-agree at the top. In each row, the first figure is the share of students who moved toward the centre of the scale and the second is the share who moved away.
2Moderately disagree
← Moved toward center
Moved away from center →
7Strongly agree
53%
6Moderately agree
47%
19%
5Slightly agree
7.6%
37%Moderate positions strengthen
4No idea / opinion
72%
the undecided develop a view
3Slightly disagree
7.5%
35%
2Moderately disagree
47%
15%
1Strongly disagree
56%
Strong opinions soften
Share of students at each starting position who moved toward or away from the center of a 7‑point scale after one conversation. Students at the poles (1 and 7) can only move inward, and matching on disagreement selects pairs that momentary error may exaggerate — but neither artifact predicts the moderates at 3 and 5 moving outward.
02 the gap between partners narrows
Nearly every pair starts with a gap — the distance between the two students’ positions on the 7‑point opinion scale. After one conversation that gap shrinks in two thirds of pairs, and it widens only in one in seven.
N = 6,392 conversation pairsMean gap 3.6 → 2.219% ended unchanged
A histogram of the change in each pair's opinion gap, from three points wider on the left to six points narrower on the right, followed by the mean gap before and after.
14% of pairs diverged67% of pairs converged
9.4%
19%
19%
20%
15%
8.4%
−3
−2
−1
0
1
2
3
4
5
6
Change in the partners’ opinion gap, in points on the 7‑point scale
Mean opinion gap
−39%
3.6
2.2
Before chat
After chat
95% CI 38%–40% · p < 0.0001
Both partners rated the topic statement before and after the conversation, so each pair yields a gap before and a gap after. The mean gap fell by 1.4 points (paired t‑test, p < 0.0001; Cohen’s d = 0.78). The 2.5% of pairs who started at the same position (Devil’s Advocate chats) had no gap to close. Matching selects pairs that start far apart, so some of this drop is regression after that selection rather than attitude change.
03 rational updating, one pair at a time
Many people expect the most common outcome to be both students moving closer to the center — but that’s not what we see in the data. More than twice as often, only one student moves toward the other. This is not compromise for the sake of getting along.
A bar chart of five outcomes for a pair, from both partners moving apart to only one moving toward the other, each with its share and its number of pairs.
The signature of rational updating
2.8%n = 177
10%n = 647
12%n = 784
22%
n = 1,438
52%
n = 3,346
Both moved apart
One moved away from partner
Neither moved
Both moved closer together
Only one moved toward partner
How student opinions changed after engaging with an opposing viewpoint. In 75% of pairs at least one partner moved toward the other; in only 2.8% did both move apart. “Only one moved toward partner” remains the most common outcome on the hardest political topics.
04 students understand opposing perspectives
After every conversation, each student answers five questions about what their chat partner actually argued — their reasons, objections, and rebuttals — and a majority of students ace it.
N = 14,473 quizzesMean 4.3 / 5 (87%)Median 5 / 5
A bar chart of the six possible quiz scores, from nought out of five to five out of five, each with its share and its number of quizzes.
The most common score is a perfect 5 / 5
0.26%n = 38
1.2%n = 180
3.3%n = 478
11%n = 1,526
29%
n = 4,137
56%
n = 8,114
0 / 5
1 / 5
2 / 5
3 / 5
4 / 5
5 / 5
Sway’s post‑chat Understanding Quiz tests each student’s understanding of their partner’s reasoning. Each quiz has five four‑choice questions, AI‑generated from the chat transcript. Results reveal students are listening closely, understanding, and remembering their partner’s reasoning in detail.
05 students treat each other with respect
One‑on‑one, students engage with respect and curiosity; the hostility familiar from social media rarely appears.
Moderation at scale
<0.46%
of student messages flagged as potentially unconstructive.
250,000+MESSAGES REVIEWED
Personal attacks and insults are blocked automatically.
Student protections
1
Encrypted conversations
Instructors and staff cannot read identifiable transcripts or view any individual student’s opinions.
2
Students control what’s shared
Students choose what goes to instructors and can delete their accounts at any time.
3
Never used to train commercial AI
Student data stays out of third‑party model training.
4
Research is opt‑in
Participation in studies is always a student’s own choice.
Across more than 6,000 ratings by students from across the United States, 92% agreed that “my partner was respectful” (1.5% disagreed) and 90% agreed that “I was not offended by my partner’s perspective” (2.6% disagreed).
06 Guide treats all students with respect
Guide, Sway’s AI discussion facilitator, participates in all student chats. After each chat, students rate the statement:
“Guide treated me and my partner with equal respect.”
Only 3% disagree.
N = 5,637 ratings
86%AGREE
12%NEUTRAL
1
Guide doesn’t offer opinions.
It joins every conversation to pose challenging questions, surface hidden assumptions, and de‑escalate tense moments — without taking sides. Instead of arguing, it pushes students to reason more carefully and engage the strongest version of the opposing view.
2
600+ politically diverse Americans found no bias.
Partisan reviewers evaluated real transcripts on abortion, healthcare, immigration, and trans athletes. Left and right alike judged Guide even-handed, and both correctly flagged bias when it was experimentally introduced.