Uploaded September 2026 | Updated November 2026, 3 hours ago
Pair two students who strongly disagree about immigration or abortion, tell them to hash it out, and most people expect it to go badly. Across almost 14,000 conversations, it doesn't.
Simon Cullen, Senior Fellow at Heterodox Academy, presents two years of data from Sway, a chat platform that pairs students who disagree about statements their instructors write and drops an AI facilitator into every conversation. Since launching in summer 2024 it has hosted nearly 14,000 one-on-one discussions across more than 100 institutions, generating almost half a million messages and 18 million words of transcript.
What that data shows has frustrated a lot of common assumptions:
Self-censorship looks like an audience problem. Over 90% of students report feeling comfortable, and that number stays flat across roughly 17 subject areas, including Israel, abortion, and race and policing.
Two thirds of pairs move closer together after chatting. Pairs begin 3.6 points apart on a seven-point scale and close nearly 40% of that gap in about 30 minutes. Simon tests whether that is real persuasion or just measurement noise, and shows why the noise explanation doesn't hold.
Students understand the other side better than expected. On a closed-book quiz about their partner's reasoning, a perfect score is the most common result.
Hostility barely appears. Under 0.5% of nearly a quarter million messages were flagged as even potentially unconstructive.
Partisans on the left and right both rated the AI facilitator as even-handed. The agreement was so uniform that the team inserted bias into transcripts to check whether people were paying attention. Everyone caught it. What drives perceived fairness is procedural: hit both sides equally hard.
And in the deepest implementation so far, the facilitator chose silence more often as the semester went on, as students needed it less. Simon flags this one as preliminary and not yet replicated.
His argument: the right role for AI in higher education is facilitative, not generative. Not a tool that thinks for students, but one that pushes them to think better.
Chapters
0:00 Welcome and introductions
1:40 What Sway is and how instructors use it
4:18 How students get matched with someone who disagrees
5:15 The understanding quiz and post-chat survey
6:50 How students actually rate the experience
7:37 Self-censorship and the audience problem
9:06 What students find on the other side
10:14 Student feedback on the AI facilitator, in their own words
13:21 The dataset: 14,000 conversations, 12,000 students
14:19 What topics instructors choose
15:27 How attitudes change after a conversation
17:30 Two thirds of pairs move closer together
18:20 Persuasion or measurement artifact?
21:16 Comprehension quiz results
23:21 Findings beyond the classroom
24:08 Why it doesn't blow up
24:49 Testing the AI facilitator for political bias
26:48 Procedural fairness is what people actually judge
27:40 The AI that learns when to stay silent
29:07 Audience discussion and Q&A
About the speaker
Simon Cullen is a Senior Fellow at Heterodox Academy and co-founder and president of Disagree Wisely. He developed the award-winning Dangerous Ideas in Science and Society course at Carnegie Mellon, was a Faculty Research Fellow at HxA's Segal Center for Academic Pluralism in 2025, and won HxA's 2024 Open Inquiry Award for his work on AI facilitation.
Links
Sway: swaybeta.ai
Sway research: swaybeta.ai/research
Event page: heterodoxacademy.org/events/facilitating-constructive-disagreement-at-scale-surprising-findings-from-14-000-one-on-one-student-conversations
Become an HxA member: heterodoxacademy.org/become-a-member
This colloquium is part of Heterodox Academy's Research Colloquium Series, a virtual seminar series highlighting scholarship from our community. Colloquia are offered to HxA members, who get live access and Q&A. Membership is free.
Great Minds Don't Always Think Alike.
#HeterodoxAcademy #ConstructiveDisagreement #AIinEducation
Pair two students who strongly disagree about immigration or abortion, tell them to hash it out, and most people expect it to go badly. Across almost 14,000 conversations, it doesn't.
Simon Cullen, Senior Fellow at Heterodox Academy, presents two years of data from Sway, a chat platform that pairs students who disagree about statements their instructors write and drops an AI facilitator into every conversation. Since launching in summer 2024 it has hosted nearly 14,000 one-on-one discussions across more than 100 institutions, generating almost half a million messages and 18 million words of transcript.
What that data shows has frustrated a lot of common assumptions:
Self-censorship looks like an audience problem. Over 90% of students report feeling comfortable, and that number stays flat across roughly 17 subject areas, including Israel, abortion, and race and policing.
Two thirds of pairs move closer together after chatting. Pairs begin 3.6 points apart on a seven-point scale and close nearly 40% of that gap in about 30 minutes. Simon tests whether that is real persuasion or just measurement noise, and shows why the noise explanation doesn't hold.
Students understand the other side better than expected. On a closed-book quiz about their partner's reasoning, a perfect score is the most common result.
Hostility barely appears. Under 0.5% of nearly a quarter million messages were flagged as even potentially unconstructive.
Partisans on the left and right both rated the AI facilitator as even-handed. The agreement was so uniform that the team inserted bias into transcripts to check whether people were paying attention. Everyone caught it. What drives perceived fairness is procedural: hit both sides equally hard.
And in the deepest implementation so far, the facilitator chose silence more often as the semester went on, as students needed it less. Simon flags this one as preliminary and not yet replicated.
His argument: the right role for AI in higher education is facilitative, not generative. Not a tool that thinks for students, but one that pushes them to think better.
Chapters
0:00 Welcome and introductions
1:40 What Sway is and how instructors use it
4:18 How students get matched with someone who disagrees
5:15 The understanding quiz and post-chat survey
6:50 How students actually rate the experience
7:37 Self-censorship and the audience problem
9:06 What students find on the other side
10:14 Student feedback on the AI facilitator, in their own words
13:21 The dataset: 14,000 conversations, 12,000 students
14:19 What topics instructors choose
15:27 How attitudes change after a conversation
17:30 Two thirds of pairs move closer together
18:20 Persuasion or measurement artifact?
21:16 Comprehension quiz results
23:21 Findings beyond the classroom
24:08 Why it doesn't blow up
24:49 Testing the AI facilitator for political bias
26:48 Procedural fairness is what people actually judge
27:40 The AI that learns when to stay silent
29:07 Audience discussion and Q&A
About the speaker
Simon Cullen is a Senior Fellow at Heterodox Academy and co-founder and president of Disagree Wisely. He developed the award-winning Dangerous Ideas in Science and Society course at Carnegie Mellon, was a Faculty Research Fellow at HxA's Segal Center for Academic Pluralism in 2025, and won HxA's 2024 Open Inquiry Award for his work on AI facilitation.
Links
Sway: swaybeta.ai
Sway research: swaybeta.ai/research
Event page: heterodoxacademy.org/events/facilitating-constructive-disagreement-at-scale-surprising-findings-from-14-000-one-on-one-student-conversations
Become an HxA member: heterodoxacademy.org/become-a-member
This colloquium is part of Heterodox Academy's Research Colloquium Series, a virtual seminar series highlighting scholarship from our community. Colloquia are offered to HxA members, who get live access and Q&A. Membership is free.
Great Minds Don't Always Think Alike.
#HeterodoxAcademy #ConstructiveDisagreement #AIinEducation










