Implementation showcase
strangers by design
Two professors linked a freshman seminar at one university to a lecture class at another, and for fourteen weeks matched their students to argue with someone they would otherwise never meet.
“I think you make some good points here about Taiwan.”A student, conceding an argument about China's military threat to the US
He had opened the conversation arguing that a conflict over Taiwan wouldn't really count as a military engagement between China and the United States — and, in the same message, that mutual deterrence was exactly what would keep such a conflict from happening. His partner, a stranger at another university four hundred miles away, caught the contradiction: deterrence only matters if the conflict could actually draw the US in. Guide put the inconsistency to him directly: if you don't think Taiwan involves the US militarily, why would deterrence apply at all? He read the question, and changed his answer.
That exchange happened in week thirteen. Two political science instructors — one running an eighteen-student freshman seminar at a private university, the other teaching a ninety-one-student class at a large public university several states over — agreed to run their courses in parallel. Once a week for fourteen weeks, students at both schools opened a Sway assignment on the same contested subject — campus speech, wealth, immigration, gender, foreign policy, artificial intelligence — and Sway matched them, wherever possible, with someone on the other campus who disagreed. Two classes, in two states, arguing weekly across a line neither would otherwise have crossed.
Neither school is named here. Instead: the seminar (small, private, eighteen students) and the lecture class (large, public, ninety-one). In excerpts below, the tag under each speaker's name marks which one they were in. No student is named either; everything in quotation marks is verbatim, and every claim about the data is one we checked against the underlying conversation and survey records ourselves.
The main read is about fifteen minutes. The + drawers hold more conversations, the full week-by-week map, and the numbers behind each claim, and roughly double it.
How this showcase was built
Sway generates a deidentified report after every assignment — the distribution of opinions on each statement, an AI-written summary of every conversation of ten or more messages, post-chat survey results, and students' optional written feedback. This showcase draws on all fourteen of the course's weekly reports, plus the aggregated report covering the full term, together with the underlying conversation and survey records.
Two things made this implementation unusual enough to check by hand rather than take on faith. First, we read several dozen full transcripts — a broad selection across all fourteen weeks, weighted toward the ones that turned out strongest — before excerpting anything. Second, we traced five students across their entire semesters using the stable identifier behind each account (the same student keeps the same account, and therefore the same underlying data, all term, even though nothing about them is exposed to instructors or disclosed here). Quotations from students and from Guide are verbatim, informal spelling and punctuation included; where a student's own message named their school or another real person, we removed or generalized it, and we have flagged every real first name we encountered so none of them appear on this page.
The five students profiled below were chosen from the course's own participation and opinion-movement records for the range of their semesters — not hand-picked for agreement with the platform. Three go to the seminar, two to the lecture class; that split is deliberate, not because the lecture class had less to offer, but because our brief for this piece was to feature range among students who found the exercise worthwhile, and a full accounting of the lecture class's experience is more mixed than the seminar's.
The design
Eighteen, and ninety-one
The seminar was small on purpose. It was a required freshman course at a private university, built around one theme — the practice, not just the theory, of disagreement — and this was the instructor's first time ever teaching first-years after a long career teaching upperclassmen and graduate students. The lecture class was a standard-sized political science course at a large public university, five times the size, meeting the material as one unit among several rather than as the whole point of the term. Neither instructor changed the other's syllabus. What they agreed to share was the weekly Sway assignment: the same statement bank, the same week, students from both schools thrown into the same matching pool.
That size difference did something to the two classes' experience of each other that the aggregate numbers make plain. Sway's matching tries to pair each student with a classmate — in this case, classmate across either campus — who holds a different opinion on the statement in front of them. With ninety-one students to draw from at the large school and eighteen at the small one, the arithmetic was never going to be symmetric.
For the eighteen students in the seminar, crossing the line was close to the default experience of the course: four out of every five conversations they had all term were with someone at the other school, a total stranger, arguing inside a different institution's politics and a different institution's version of the news. For the ninety-one students in the lecture class, the seminar's small size made the reverse true — most weeks, most of their matches were with each other, and a conversation with someone from the other school was the exception rather than the rule. Both experiences are on the record here. The seminar spent a semester in near-constant contact with strangers; the lecture class spent a semester mostly arguing with people they already half-knew, with occasional conversations reaching four hundred miles further out.
What the seminar's instructor said about it
An interview with the instructor of the small seminar is on the record for this course, conducted after the semester ended, and it fills in texture the data alone can't. Some of what she described, in her own words, deidentified here as everywhere else on this page:
Every student in her class completed at least ten of the fourteen assignments — the required minimum — and most did nearly all fourteen; a required-course completion rate she called unusual in her long teaching career ("I don't usually have assignments where 100% of people follow through and finish"). Students, most of them arriving from high schools that had, in her account, actively discouraged the airing of disagreement, told her they felt able to say things on Sway — and increasingly in her classroom too, as the semester went on — that they would not otherwise have risked in front of their peers. She banned phones and laptops in her own class sessions and built a companion assignment where each student had to research and present the strongest arguments on both sides of that week's statement before the class discussed it, so that no one arrived "intimidated," in her word, by a debate they hadn't prepared for.
Asked about trusting an AI facilitator with unsupervised student discussion — the anxiety she said most instructors raise first — her answer was that the students trusted it more than they trusted her: "they know that Guide is AI and they find it not at all partisan, not at all biased — that was never an issue," in a way that she, a visibly opinionated professor of political science, felt she could never quite achieve. She used the weekly reports in class, once building an assignment where students mapped the arguments from their own Sway conversation onto a shared diagram. She called it her favorite course in ten years of teaching. At the end of the term, she said, the students brought her flowers.
The design
Fourteen weeks, in order
The two instructors did not ease into it. The very first week put students on the record about censorship and extremism on their own social media feeds; by the second month the class was arguing abortion, gender, and Gaza. Below is the full run, in order, with one line on what the record shows from each week. Six of these weeks — marked → read the chapter — get a full treatment further down the page; the rest are summarized here from the same underlying reports and thread data.
-
1
Aug 29 · Social media
Ten statements, from censoring extremist views to whether social media has made people more politically informed.
Two independent pairs, working separately, reached the same distinction on their own — that being "politically updated" isn't the same as being "politically informed." Read the chapter →
-
2
Sept 5 · Campus speech
Administrative overreach, hate speech, punishing professors for pronoun mistakes, and which speakers a student group can fund.
Thirteen of the term's cross-campus conversations landed here — more than any other week — including the thread this page returns to more than once. Read the chapter →
-
3
Sept 12 · Conspiracy theories
The Epstein files, GMO safety, government surveillance, media bias, 9/11, vaccines and autism.
The week's honest low point: two of the September 11th conversations drifted from engineering claims into ethnic generalizations about who controls institutions. The week's high point, a GMO-safety pair, is in the "more" drawer below.
-
4
Sept 19 · Climate change
Whether people and animals will simply adapt, whether bad climate news is demotivating, who should pay.
Six threads stand out this week, four of them crossing the campus line. Read the chapter →
-
5
Sept 26 · Wealth inequality
Inheritance, flat versus progressive taxation, legacy admissions, universal basic income.
On inheritance, several students who opposed any estate tax as "double taxation" reversed once a partner explained stepped-up basis — that unrealized capital gains can pass to heirs having never been taxed at all. More than one landed on the same fix: not a principled line, but a multi-million-dollar threshold.
-
6
Oct 3 · Abortion
Whether it's murder, mandatory waiting periods, a federal guarantee of access, fathers' say, the morning-after pill by mail.
A week with a striking number of on-the-record reversals — several students narrating, in their own words, exactly which sentence changed their mind. Read the chapter →
-
7
Oct 14 · Social justice movements
Whether protest ever accomplishes anything, punishing campus protesters, the motives of people who march.
Guide's sharpest move here was evidentiary: one student built a whole case for why protest doesn't work on the strength of their own social-media feed, and was asked how representative a sample that really was.
-
8
Oct 24 · Immigration
Birthright citizenship, welfare eligibility, whether immigration depresses American wages, favoring workers who'll take jobs Americans won't.
The week's sharpest legal correction: a student's claim that a hiring preference for immigrants was unconstitutional collapsed the moment a partner asked which provision it violated.
-
9
Oct 31 · Pornography
AI-generated child sexual abuse material, age verification, whether creating explicit content should cost you your job.
One thread turned on a single move: separating a moral verdict on pornography from a factual claim about how employers and the public actually read a public image, and arguing only the second.
-
10
Nov 7 · Group fairness
Whether Black Americans have an unfair hiring advantage, coastal elites, credentialism, whether men or women have the edge in hiring today.
A student asked for evidence behind a "biology, not bias" claim came back, unprompted, with named cross-country research literature — and then, pressed further, with the specific developmental-psychology findings behind it.
-
11
Nov 14 · Sex and gender identity
Gender-affirming surgery for minors, gender-neutral restrooms, insurance coverage for adult transition care.
Restroom-access pairs converged more than they clashed: students who opened on opposite sides mostly settled on keeping gendered multi-stall facilities while adding single-stall options, an outcome neither side started with. What never resolved was whether transition surgery treats gender dysphoria or, as some students argued, merely accommodates it — a distinction the report says evidence alone couldn't settle.
-
12
Nov 21 · Non-military foreign policy
American prestige abroad, aid conditioned on politics, whether the US should try to spread democracy.
A running theme: students kept separating the ideal of democracy promotion from the question of who should be the one doing the promoting — agreeing on the first, splitting hard on the second.
-
13
Dec 2 · Military foreign policy
Whether China is a military threat, defending democracies under attack, Israel's military campaign in Gaza, US support for Israel.
The week this page opened on. Also the week two students, on their own campus, talked each other from "keep sending weapons" to "end military support, keep the humanitarian aid" in real time. Read the chapter →
-
14
Dec 9 · Artificial intelligence
AI and human creativity, AI and unemployment, AI bias, whether AI will make it unnecessary to learn to write.
The closing week, and a fitting one: an AI-facilitated argument about whether AI erodes the very capacities it assists, with neither student mentioning the setting. Read the chapter →
More from Conspiracy Theories, week 3
GMO safety, argued as a research problem
A seminar student opened by admitting she wasn't well informed on the science — and Guide called the admission out directly, asking how her confidence in dismissing the safety concerns squared with that. She went and did the reading mid-conversation, came back with the FDA's consumption-safety position and an alternative explanation for rising chronic disease rates, and ended up moving her own partner toward agreement rather than the other way around. Her own close: “This was a great discussion on GMO foods. I'm planning to research in the future on how GMOs might have harmed people in the past so I can be fully informed.”
Week 1 · Social media
The first week, and already strangers
The course opened on a subject every student already had a private, half-examined opinion about. Nothing had been built up yet between the two campuses — no rapport, no shared history of a semester's disagreements, just two students and Guide.
One thread paired a seminar student who wanted flagging over deletion against a lecture-class student defending free expression on First Amendment grounds. Neither one folded, but neither one talked past the other either:
Fringe movements gain traction by targeting online groups that may not have the resources to educate themselves. That's particularly dangerous on social media websites, where algorithms recommend content with longer retention times. Put simply, my argument is that fringe movements and extreme ideas can easily grow on social media (ex. Pizzagate) without any regulation.
I think you have a good point about algorithms often pushing more extreme/controversial content... However, censorship has historically been a tool used to undermine the perspectives of individuals in minority groups whose experiences were seen as too controversial or unpalatable to be widely spread. (An example that comes to mind is many schools' practice of banning books…)
Verbatim, trimmed where marked. The two spent the rest of the thread distinguishing "banning books" from "flagging fake news" — a line the seminar student drew and the lecture-class student, cautiously, accepted.
A separate pair, in a different thread that same week — also across the campus line — spent eight messages on whether social media has made citizens more politically informed, and reached, independently, the same distinction: that being politically updated is not the same as being politically informed. That distinction, or some version of it, recurs across most of the class's discussions that week; it wasn't limited to these two threads.
You both seem to agree that social media can effectively share basic information about events. But does this make citizens more “politically informed” in a meaningful way?
I mean I wouldn't say it makes them more "politically informed" in a meaningful way, rather just makes them more politically updated to the point where they can begin to form their own opinions
Verbatim. The two closed warmly: "I also appreciate how you were willing to try and understand my perspective... I did actually enjoy having this conversation with you, so thank you!"
The week's most substantial thread was a straight geopolitical argument that produced a real, on-the-record change of mind. Two students argued whether TikTok should be banned unless sold to an American company — one citing Chinese Communist Party data access as a unique national-security threat, the other arguing every platform has the same privacy problem and banning one company on the basis of ownership alone sets a dangerous precedent. Four days in, the lecture-class student went and checked the specifics for herself:
After looking deeper into the matter about privacy violations with Tik Tok, my stance has shifted. I was unaware how serious of an issue this was with the Chinese company. However, I do still believe that there would also be privacy violations if selling the company to America. But, this would ensure our data isn't used against us in the future.
Verbatim. She had opened the conversation firmly opposed to a ban; her recorded post-chat opinion moved two full points on the seven-point scale.
More from week 1
Flagging as coaching, not censorship
A different pair, arguing the same censorship statement, converged on fact-checking as a middle path — then had to reckon with who gets trusted to run it. "I am wary of putting faith in social media companies (which have a long history of prioritizing profits and engagement over users' well being) to implement this censorship ethically and effectively," the lecture-class student wrote, and her seminar partner agreed without pretending the problem was solved.
Numbers nobody checked
An unsourced figure — that roughly two-thirds of content creators never fact-check what they post, attributed to UNESCO — circulated through more than one thread and helped settle at least one argument on its own. In a separate thread, two students treated the rise of depression rates alongside the rise of social media as proof the first was caused by the second, without either checking whether the timing could be coincidence. Both are exactly the kind of confident-sounding claim worth correcting the following week in class, and the report flagged them for that purpose.
Week 2 · Campus speech
“I have been chosen to argue the other side”
More cross-campus conversations landed on this week's statements than on any other week of the term — thirteen in all. Two students, one from each school, were handed a statement about whether student groups should be allowed to spend university funds on speakers who might offend other students. Neither one treated it as abstract.
The lecture-class student opened by describing a controversy on her own campus — a state law restricting what public universities could fund, and a speaker invited to argue that transgender women shouldn't compete in women's collegiate sports, a topic she found genuinely harmful to a small and already-targeted group of her classmates. The seminar student raised a case from her own school: a pro-Palestinian student group that had lost its official recognition after a controversy over a literary festival, on grounds she wasn't sure she trusted.
My concern is who gets to determine who can and cannot speak on campus. There are always going to be people who are offended by someone expressing their views... I don't think student groups should be allowed to invite speakers who are like straight up Nazis or part of the KKK, but it gets more difficult when determining things like a pro-Palestine speaker who may be labeled a terrorist by pro-Israeli groups on campus. I'm wondering what a policy would look like that could make that distinction, while allowing productive discourse to exist on campus.
You seem to accept that some speakers should be excluded... but you're worried about where to draw the line. If you're willing to exclude some speakers based on their potential to harm or offend, how is your position fundamentally different from your partner's? You're both essentially arguing for content-based restrictions — you're just disagreeing about where the boundary should be.
Verbatim, trimmed where marked; the university names in the original have been generalized to "her own school."
Rather than dig in, the lecture-class student stopped and stated her actual position — on the record, mid-assignment:
I honestly struggle with the way Sway works because I can slightly agree with this topic because I believe that universities should have jurisdiction to protect their students but also completely agree with you. Censorship at [my school] specifically is rampant right now and I don't agree with it. I think that we honestly share a very similar opinion but I have been chosen to argue the other side.
Hold on. You just said you "completely agree" with your partner, and that you're only arguing your side because you were "chosen to argue the other side." But look at your actual reasoning: you supported restricting a speaker because it "opens negative conversations" about a marginalized group. You said universities should "prohibit hate speech" and have the "power" to pull funding. Those aren't throwaway positions — they reflect real values about campus speech. What do you actually believe?
Verbatim, school name generalized. She did not retreat from the question — the next message is her actually answering it.
The two spent the rest of the thread designing an actual mechanism together — a review board of students and faculty, a supermajority threshold to rescind a speaking invitation, transparency requirements for its decisions — each one testing the other's proposal against the exact case that had bothered them going in. The thread ends on a question from Guide that neither student answered: you both criticized your own schools for shutting down student expression — yet your system could produce the exact same result. How would your jury avoid the same mistakes you're criticizing universities for making right now?
Partway through the same thread, the lecture-class student misnamed the seminar student's school; the seminar student corrected her, and she apologized. We can't quote the exchange without naming a school, but it's a small reminder that the two of them were, among other things, working out who exactly they were talking to.
More from week 2
“In my three years here”
A pair debating whether professors should excuse students from sensitive class discussions moved from an abstract "student comfort versus academic freedom" framing to something more grounded, once the lecture-class student offered her own record: "in my 3 years [here], I have never had a class where students weren't free to come and go as they please. We are all adults." Her seminar partner, pushed by Guide to test that against a critical-race-theory example, ended the thread conceding: "After our conversation, I would probably say that I mildly agree with the proposition that students should be allowed to leave if there is content that makes them uncomfortable."
Hate speech, and where the conversation almost ended too soon
Arguing whether hate speech should be illegal, a seminar student worked through the distinction between speech that's already illegal (threats, incitement, defamation) and speech that's merely offensive — and her lecture-class partner held the free-expression line right up until she tried to sign off. Guide caught it: "Hold on — you're ending the conversation right when you two were getting to the heart of your disagreement!" She stayed, and closed by conceding that "educational oversight, when applied well and proven effective, can be an alternative to criminal law."
Punishing professors for pronoun mistakes
A pair spent ten exchanges trying to draw a workable line between an honest mistake and a malicious one, converging on training plus escalating consequences for repeat offenses — then had to work out, live, whether the same standard should apply to a professor versus a high school teacher, catching and correcting the mix-up themselves mid-thread.
Mascots, and who gets to decide
Two lecture-class students spent five exchanges on Native American sports mascots and landed on a workable distinction — tribe-specific names like "Seminoles," which the affected tribe can be consulted on directly, versus generic names like "Braves," which would require consensus across many tribal nations to change. Guide's closing question stayed open: does requiring consensus from hundreds of distinct tribes make change practically impossible?
Week 4 · Climate change
What "adapt" was being asked to mean
Six of this week's threads stand out, four of them crossing the campus line. The statements weren't softer than the rest of the term's; if anything, "climates have always changed, so people and animals will adapt as they always have" turned out to be one of the harder sentences the class was handed all semester, because it hides a philosophical question inside what reads like a factual one.
One thread paired a seminar student who disagreed against a lecture-class student who agreed. He opened by taking the statement's premise seriously and pressing on what "adaptation" is even claiming:
I am arguing that individual species may not survive, but life overall still persists. Extinction is a natural process of history... Species that cannot adapt to the new environment will die out, and those that are capable will flourish.
Specifically as a result of climate change and human impacts on the environment, since 1970, 60% of animal populations have been wiped out, which isn't a "natural process" when it happens at the rate it does and because of human-made changes to environments. Adaptation to me doesn't mean extinction, it means species changing over time to continue surviving in their environment.
Verbatim. She went on to catch a factual slip — her partner had cited deforestation to explain coral-reef and Arctic species loss, environments with no forests to clear — and he conceded the point directly rather than defend it.
He held onto a narrower claim — that human technology, unlike biological adaptation, might still let people keep pace — and she went after that too, turning his own example against him:
I don't believe that humans, with the way we're operating now, are currently successfully adapting to climate change. Our most significant new technology is generative AI which quite literally destroys the environment with every output.
You brought up generative a.i., which indeed does cause climate change. However, a single commercial airline flight produces significantly more CO2 output than thousands of AI models combined. I feel that A.I. can actually help humans adapt to climate change if utilized correctly, as shown in disaster prediction systems and agricultural optimization.
Verbatim. The thread closed with her arguing for legislation and renewable investment rather than individual-scale fixes — a position neither student had opened with.
A second cross-campus thread on the same statement pushed the philosophical question to its limit: if adaptation, taken literally, allows for mass death as long as some population survives, does that still count as adaptation working? Guide pushed both students past an easy agreement toward the actual disagreement underneath it, closing on a question neither answered: will adaptation work this time like it has before? And a third, on a related statement about doom-laden climate coverage, produced one of the week's cleanest unresolved endings — Guide's who decides what the "right mix" is? left standing after both students admitted they didn't fully know.
More from week 4
Whether environments themselves can adapt
Two lecture-class students, not matched across the campus line, argued whether environments themselves can be said to "adapt," or only the populations within them. One brought a professor's framing from another class into the chat verbatim: "climate change is real; we caused it; we can fix it." They ended agreeing that mass death through natural selection technically satisfies the original statement — and one of them explicitly registered how strange it felt to arrive there: "It's a harsh reality, but it's just the nature of how the world works."
Electric vehicles, argued as policy design
A same-campus pair moved from "should the government subsidize EVs" to a genuinely granular argument about where in the supply chain an incentive should land — consumers, manufacturers, or the mining and recycling of battery materials — and closed debating whether a mandate, not just an incentive, might eventually be necessary.
Week 6 · Abortion
Reversals, on the record
By the sixth week, the two classes had a method. A striking number of students explicitly revised their recorded position this week, and several of them wrote down exactly what changed it.
One thread put a seminar student who supported mandatory waiting periods against a lecture-class student who called the policy singling out women's medical decisions for a scrutiny no other procedure receives. Guide gave the seminar student almost no room to retreat into vagueness:
You just said the "possibility of becoming a person" makes abortion different. But here's the thing: if the waiting period is justified because abortion involves ending a potential life, then you're essentially saying the fetus has some moral status that makes this decision different from other medical choices. Is that your actual view? And if so, does a waiting period actually address that moral concern — or does it just delay that same outcome?
SPECIFICALLY, my very real argument is that abortions are very emotional, moral decision;. other medical conditions are only effecting one life, your life. But, abortions are ending another POSSIBLE life. Waiting period doesn't question the ability of women to make decisions, but it EMPOWERS them to make a socially and emotionally informed decision
Verbatim, capitals his own. He held that line for several more exchanges before the thread's real turn.
What actually moved him wasn't a counterargument so much as being made to price his own proposal. He suggested a compromise — a psychological evaluation before an abortion, waived if the patient passed — and Guide priced it immediately: who gets to define "true informed decision making" for women seeking abortions? If a woman schedules an appointment, researches her options, and shows up ready to proceed — why does she need to prove anything to a psychologist? By the end of the thread he had narrowed his own proposal from a mandatory delay to an optional, voluntary conversation — which Guide named plainly as what it was: you've basically abandoned your original position that waiting periods should be required.
A second thread that week produced a rarer thing — a student narrating her own mind changing, in writing, on the assignment record. Arguing whether abortion's moral status should be left to individual conscience rather than law, a lecture-class student built a dense, evidence-heavy case for the pro-choice reading of viability and consciousness, and her seminar partner, who had opened undecided between positions, wrote back:
I think when it comes to pro choices and lifers debating, it usually comes down to the immorality of "killing" a fetus... It is sentience and viability (the ability to live outside of the womb) that gives fetuses moral value, and consciousness does not develop in fetuses until about 20-24 weeks. 99% of abortions happens LONG before this… I'd love to know if this has affected your position on the issue at all, let me know!
I changed my answer to the question to moderately disagree. There's some wiggle room left on account of when people's beliefs are based on personal interest or religion/faith, then that should not be something taken into consideration when writing laws. Thats just being nitpicky though, overall, I do think a basic level of protection has to be reinstated.
Verbatim, trimmed where marked. Her recorded pre-chat opinion had been genuinely uncommitted; her post-chat answer, and her own account of why, are both on the record.
More from week 6
A federal guarantee, argued through the Civil Rights Act
A lecture-class student defended a federal right to abortion access by analogy to federal civil rights protections that limited state power for a clear public good; her seminar partner, who opened firmly against a federal guarantee on constitutional grounds, ended the thread genuinely undecided: "At this point, I am unsure of whether I support a federal law guaranteeing access to abortion." Neither treated that as a loss.
Two students who kept the disagreement and the respect both
A pro-choice seminar student and a pro-life lecture-class student spent five exchanges on whether a federal law would diminish the value of fetal life, each taking the other's strongest point seriously without conceding the underlying disagreement. It closed on shared ground neither had opened with — better healthcare, childcare, and family-support systems — with the pro-life student's closing line: "True compassion doesnt mean ending one life to make another's circumstances easier, it means improving the circumstances so that both can live."
Week 13 · Military foreign policy
China, Taiwan, and Gaza
This is the week this page opened on. Its statements included whether China poses a military threat to the United States, and whether Israel should continue its campaign in Gaza until Hamas is eliminated.
The thread this page opened with belongs here. A lecture-class student argued for the statement by naming specifics — China's carrier fleet, its stealth aircraft program, and above all Taiwan, where a US commander had publicly floated the possibility of American forces playing a role in any Chinese invasion. His seminar partner leaned entirely on mutual assured destruction as a backstop, until Guide pointed out the hole in that:
There's a tension in your partner's position here worth exploring. He says he doesn't think a conflict in Taiwan counts as military engagement with the US — yet he also invokes mutually assured destruction as a deterrent for Taiwan scenarios. But MAD only works if both sides believe the conflict could escalate to direct US-China war. If he doesn't think Taiwan involves the US militarily, why would MAD apply at all? Can you press him on this inconsistency?
Why do you claim that the conflict in Taiwan isn't considered a military engagement with the U.S., yet insist on MAD as a deterrent against the Taiwanese situation?… if you confirm the conditions you've presented, the logic that deterrence against the Taiwanese situation is a means of preventing conflict between the US and China, even though there is no conflict between the two, is undeniably flawed.
Verbatim, trimmed where marked; partner's name removed.
The seminar student conceded outright, on the record: “I think you make some good points here about Taiwan. Honestly when I first read this question I was only really thinking about direct conflict between the U.S. and China. Now I can see a bit better how Taiwan and other similar situations could escalate into conflict between these superpowers.” His partner's closing line: “I appreciate your articulate delivery of your points, and I certainly see your logic as correct.”
A seminar student who had agreed with the statement was matched with another seminar student, not across the campus line this time. Rather than rebut him, she asked him what he thought Israel should actually do next:
Provided how the war has progressed, what do you believe is the right course of action that either Israel or Palestine should take?
i think the right course of action that mainly israel should take is meeting the urgent humanitarian needs and holding themselves accountable to what they have done. i think this war has escalated to a level that israel needs to also be held accountable globally and not dismissed.
You've shifted significantly here. Earlier you argued Israel should continue until Hamas is eliminated. Now you're emphasizing Israeli accountability and criticizing U.S. support. How do you reconcile these positions? Can Israel both "continue its military campaign until Hamas is eliminated" and be condemned for the harm it's causing? What does accountability look like while the campaign continues?
no, actually i would like to stand with my shift here. israel needs to be help accountable and should not be the one controlling others or saying when it ends and on only their terms.
Verbatim. He went further, unprompted, criticizing US political support for the campaign in his next message.
A separate, same-campus thread that same week moved in the same direction, on the narrower question of US military aid to Israel: a lecture-class student who opened defending continued support ended the thread writing “I feel like my opinion has shifted towards no providing military support to Israel,” after her partner reframed the disagreement around civilian protection rather than alliance loyalty. Her partner: “I appreciate you being open to shifting your view… I feel like this conversation has been really good.”
Week 14 · Artificial intelligence
The last week: arguing about AI itself
The last week of the semester of AI-facilitated arguments was spent arguing about AI. One thread is worth noting for where it happened as much as what was said — two students debating whether artificial intelligence will meaningfully accelerate medical progress, inside an AI-facilitated conversation, without either of them remarking on the fact.
AI is a force multiplier for physical research. Medicine depends on choosing the right experiment out of thousands... When it comes to speed and accuracy, I think AI is inevitable for saving more lives.
I think AI, particularly machine learning, identifies statistical correlations in data rather than causal biological mechanisms. Medicine depends on causation not correlation... AI refines existing knowledge; it doesn't generate fundamentally new medical understanding.
Verbatim, trimmed where marked. Neither backed down over the following six exchanges, but both narrowed their claims until the real disagreement was legible.
What the two of them actually converged on, without either fully conceding, was a definition of the word "extensive" — whether cumulative, marginal efficiency gains inside the drug-discovery pipeline count as an extensive medical advance, or whether the word requires something closer to a systemic shift in what actually kills people. The seminar student's closing distinction was sharp enough to be worth keeping on its own: “AI helps the pipleine run better it doesnt redefine what ultiamtely limits survivial.”
Elsewhere in the week, a debate over whether AI will worsen inequality produced one of the term's more pointed exchanges on a subject the rest of the syllabus had circled all semester — who bears the environmental cost of the technology everyone was, in that exact moment, using. A lecture-class student raised data-center siting directly:
AI increases inequality because of environmental racism. Data centers used to train AI are disproportionally located in black and indigenous regions of the country. These data centers are terrible for the environment… These minority communities are targeted by the placements of these data centers and have to live with the consequences.
I agree that data centers do cause high levels of pollution and I'd advocate for legislation that prohibits the building of these centers in marganizlaized areas.
Verbatim. The two spent the rest of the thread disagreeing, respectfully and at length, over whether AI-assisted "innovation from home" offsets the jobs it displaces.
More from week 14
“Empirically, it's just false”
Arguing whether AI's resource demands will cause environmental disaster, a seminar student and a lecture-class student worked through water usage, energy projections, and the honest admission that neither of them could actually predict the technology's trajectory decades out — and ended agreeing they'd basically agreed all along, the seminar student conceding: "This question is hard too, cause empirically (as of now), it's just false."
Bias, traced to its source
A pair arguing whether AI will amplify hiring and criminal-justice bias worked all the way down to proxy variables — whether stripping race from a dataset actually removes racial bias, or just hides it inside zip codes and names. Neither had a clean answer by the end, and the seminar student said so plainly: "I don't know the way forward for this specific issue."
The students
Five students, two campuses
The chapters above follow the weeks. Sway also lets a single student's whole semester be read as one story — the same account, the same underlying data, from the first week to the last, even though nothing about a student's identity is ever exposed to their instructor or disclosed here. We traced five: three from the seminar, two from the lecture class, chosen from the course's own participation and opinion records for the range of their terms, not for agreement with the platform. The names below are pseudonyms.
Each row of fourteen dots marks the fourteen weekly assignments in order. A solid dot means that week's partner was on the other campus; a pale dot means the partner was on the student's own campus; no dot means the student sat that week out.
Miriam · seminar
Miriam's semester ran cross-campus almost every week — nine of her eleven conversations crossed the line, including two with the same lecture-class partner on separate occasions. Her pattern, week after week, was to grant her opponent's strongest point and then relocate the disagreement to a narrower one she thought actually carried the weight. Arguing Native American mascots, she granted that unanimous tribal consent was unobtainable, then shifted the question to whether an elected tribal body's objection should be dispositive on its own. Arguing a hiring preference for immigrants willing to take jobs Americans won't, she absorbed her partner's evidence that recent labor protections were real and working, then narrowed her objection to the single word "favor." She is also the rare student in this record who went and did homework mid-argument: challenged on a GMO-safety claim she'd made without much basis, she came back citing the FDA's actual position and ended up moving her own partner, not the reverse. Her biggest recorded shift came the same week — arguing against censoring extremist content, she and her lecture-class partner built a shared position, fact-checking over deletion, that neither of them arrived with, and her own opinion crossed the scale's midpoint by the end.
“Imagine if someone took your cultural heritage and stuck it on a sports team without your consent. And now, they call it honor?”
Owen · seminar
Owen is the student this page opened with. He argues from mechanism, not from stated values — how insurance categorizes a diagnosis, what a legal standard actually requires, whether a policy proposal survives its own logic — and that style has a specific failure mode and a specific strength. The failure mode showed up the same week he argued that insurance shouldn't be required to cover adult transition surgery: pressed on why insurers already cover mental-health conditions people don't choose, he never conceded a word in the transcript, and his recorded opinion nonetheless collapsed five points by the end, the machinery of the argument apparently outlasting his actual commitment to it. The strength showed up on Gaza, where a partner didn't attack his morality but the coherence of his war aim — "how would anyone verify Hamas has been eliminated?" — and he had no way to absorb that and stay where he was. He is not easily talked out of a position, but he also does not hold one once its logic has failed, and he does not pretend otherwise when that happens.
“I think there shouldn't be ‘comfort’ in these campuses. The point of being surrounded by diversity is to explore things that we are not use to.”
Callum · seminar
Callum crossed the campus line more than almost anyone else in the seminar — twelve of his thirteen weeks — and the reader will already have met him twice above, in the climate chapter (the "60% of animal populations" thread) and the campus-speech chapter (the review-board thread). He is one of the sharpest reasoners in the entire cohort, seminar or lecture class, even though his own satisfaction ratings for those same conversations were only ever average for his campus — a split between the quality of his arguments and how much he seems to have enjoyed making them. His signature move is almost never arguing the statement he was handed; he argues what it means. Told immigrants shouldn't receive welfare benefits, he didn't dispute the premise so much as insist on a definition: "social welfare is meant for those who need it the most, not whoever pays the most for it or else it's not a social safety net, it's a private insurance system." His two biggest recorded shifts both moved him further from the statement he'd been assigned, not toward it — a reasoner who uses an argument to sharpen a position, not test whether to abandon it.
“I never said that nazis or the KKK should be allowed to speak because the line is hard to define, just that the fact is that the line is hard to define.”
Bea · lecture class
Bea's twelve threads are consistently sharp, and reading them shows why. Her characteristic move is to take a claim her partner offered in their own support and show it actually cuts the other way — told that deterrence proves China isn't a real threat, she pointed out that deterrence is precisely what a real threat produces: "Deterrence is not something that states need against harmless actors." Told an AI-generated skill's disappearance would be "damaging," she noted the word itself gave away the argument: "if AI were truly making writing unnecessary, then losing those skills wouldn't be a problem." When Guide pushed her to back up a claim about biological sex differences in career choice with actual evidence, she came back citing named cross-national research; pushed again, she came back with the underlying developmental-psychology findings. Her four cross-campus conversations included her strongest thread of the term and her single largest recorded change of mind — a real-time reversal on the TikTok statement, mid-thread, after she went and checked the facts herself.
“The problem I see with your argument is that you are using deterrence as proof that China isn't a threat. However deterrence only exists because China is a threat.”
Marcus · lecture class
Marcus is the only student in this sample who did all fourteen weeks. An international student arguing in his second language, he was consistently the most prolific writer in almost every conversation he had — often outwriting his partner three or four to one — and he is the only student in this record who directly coaches his own partners' effort mid-thread. His two conversations with seminar partners were, on the numbers, his two strongest of the semester: on China, a seminar partner conceded to him outright after Guide handed Marcus the exact contradiction to press; on abortion, Marcus's own recorded position moved four full points over the course of the conversation, in the exchange right after his partner asked him a single pointed question — "why even bother getting pregnant if you know that you are sick" — and he narrowed a broad claim to a narrow one in his very next reply: "the only case I mention… is when a pregnant woman in an abnormal condition has an unintended pregnancy." Whatever he was rating so consistently high, it wasn't always his own accuracy — one of his sharpest tones came defending a claim he couldn't actually source. But arguing in a second language, against strangers four states away, held his attention for fourteen straight weeks.
“It needs logical explains and long reasons&evidences for progress bar goes up. Keep in mind this!! Progress bar does not go up only one to two short answers.”
Three of these students crossed the campus line almost every week; two crossed it rarely, but their cross-campus conversations were consistently among their strongest. None of the five converted to their partners' politics; all five left a semester-long record of arguments getting sharper, narrower, and more precisely stated than where they started.
The measures
What moved
Sway records a student's opinion before a conversation and again after. Across the semester, 642 matched pairs were recorded on the statements students discussed. This is what they show:
642 matched pre/post opinion pairs across fourteen weeks, on a 1–7 scale. "Toward the opposing side" means the student's post-chat opinion sat closer to (or across) the scale's midpoint than where they started.
The honest caveat, stated the way the last two showcases stated it: opinions were measured immediately after each conversation, and this course had no control group. 21% of matched pairs moved away from the midpoint — a real number, not a rounding error, and higher than the comparable figure from this site's other full-semester showcase (14%). Some of that is plausibly the subject matter: this class spent fourteen weeks on statements built to be genuinely contestable, week after week, rather than easing in and out with softer material. Movement ran in both directions throughout, including students who left a conversation more convinced of where they started. Sway takes no side on which direction is correct; revision, not conversion, is the outcome it is built to produce.
The students' own assessment of the conversations, item by item, is less ambiguous:
I felt comfortable sharing my honest opinions with my partner
n = 304 · 94% agreed, across statements ranging from campus mascots to Gaza
It was valuable to chat with a partner who did NOT share my perspective
n = 307 · 77% agreed
My partner had better reasons for their views than I expected
n = 308 · 59% agreed — an admission of intellectual humility, on the record
Guide treated both partners with equal respect
n = 308 · 88% agreed, on statements about Gaza, gender, and race
The rest of the survey, including the softer numbers
Across 649 post-chat survey responses: 74% rated their chat "Good" or "Awesome," 7% negatively. 70% said Guide contributed the right amount to the conversation. On the single item that asked most directly whether a given conversation changed a student's mind on the spot — a smaller, less-often-answered item than the ones above — the answer was genuinely mixed: 32% agreed, 37% disagreed, and the rest were neutral. It's the softest number on this page. Read next to the opinion-shift chart above, it's also a reminder that "did this feel like it changed my mind" and "did my recorded opinion actually move" are different questions with different answers — the same pattern the site's two other implementation showcases report. The conversations, not either chart alone, are the evidence.
These figures track Sway's cross-institutional numbers on every item — see the full survey data from 11,000+ students — and this course's comfort and equal-respect numbers run a few points above the platform average.
The students
In their own words
Post-chat feedback is optional and free-form; nearly two hundred comments came back over the semester, from both campuses. A selection, verbatim:
It was very comforting to have a nice back and forth with someone despite our differences. It gave me confidence to try to find common ground with people I may think I disagree with. We likely have more in common than I'd realize!
It was much more civil, thought out, and logical! It felt like I was talking to another well-reasoned person rather than shouting against a brick-wall lol.
Sway, being fully online and at our own pace, allows me to think more deeply about my answers before I say them, allowing me to give the best answer possible. In class discussions, I cannot do this because I have to respond almost immediately.
Yes, as a bigger writer than a speaker, I feel more confident in participating in classroom discussions… this assignment layout gives me the opportunity to be able to write down my thoughts to understand them clearly.
I am more confident in expressing my views and am not afraid about others labeling my views as hostile or dangerous.
This chat helped me learn how to work with someone else to come up with solutions to issues. This is a skill that I hope to develop further… the chat was very fun and informative.
I am normally very hesitant to trust LLMs, especially for assignment completion, but I think Guide is truly an objectively useful tool.
I changed my answer to the question to moderately disagree… overall, I do think a basic level of protection has to be reinstated.
Including the criticism
The feedback was not uniformly warm, and the real frustration in it is concentrated more heavily among the lecture class than the seminar. Some of it names problems worth naming: students who felt an assigned position didn't match their actual view, students who found a 30-minute minimum conversation length arduous on a busy week, students skeptical that a chat window replicates in-person nuance. A smaller number, mostly in the large class, described the assignment as a chore that added little to a subject they felt equipped to discuss without it. One line, representative of the sharper end: "I don't need a little yes-man trying to chime in about the points that I am making when I am in a discussion." The seminar's near-universal completion rate and its instructor's account of the semester sit at one end of this course's range of experience; comments like that one sit at the other.
One recurring complaint is worth naming and setting aside on purpose: several students that semester described the assignment's progress indicator getting stuck near completion despite an active, substantive conversation continuing underneath it. That was a real bug in how progress was calculated at the time, and it has since been rebuilt; student reviews of the newer version are markedly better on exactly this point. We mention it here only because the record should be accurate, not because it reflects the platform as it exists today.
The instructors
What came back to the two classrooms
Neither instructor could read a single transcript directly — conversations are private, and student names are stripped from everything Sway reports. What came back after each of the fourteen weeks was the report this page is built from: which arguments recurred, where pairs found common ground, which disagreements never closed, and which confident-sounding claims spread through a class without anyone checking them — an unsourced figure in week one claiming most content creators never fact-check their posts, the drift from engineering claims into ethnic generalization in week three's September 11th threads, unsourced donor-access and hate-crime figures that went unchallenged in the social-justice-movements week. Those are exactly the kind of finding an instructor can correct in the following week's lecture, and both instructors' reports show them doing it.
The seminar's instructor, in her own account, used the reports for more than correction. She built a class exercise around them, having students map the arguments from their own Sway conversation onto a shared diagram — a way, she said, of turning a private chat into a shared, structured artifact the whole class could learn from together. And she described something that surprised her about trusting an AI, rather than herself, to referee her students' hardest arguments: it wasn't a liability. Students, she said, trusted an AI facilitator's neutrality in a way they were never quite able to extend to a visibly opinionated professor of political science — "they know that Guide is AI and they find it not at all partisan, not at all biased. That was never an issue." For a freshman seminar built entirely around the practice of disagreement, an unbiased referee was close to the whole design, not an incidental feature of it.
Each week's report also drafts a piece of the following week's teaching, unprompted. Its "for your next class" section turns a course's own unresolved arguments into ready-made exercises — a re-vote on hate speech paired with an argue-the-other-side exercise, a think-pair-share on who should decide when the group a rule protects doesn't agree among itself, a writing prompt built from a case where a scientist's politics got treated as evidence about the quality of their research. And every week closes with a recap written in Guide's own voice, addressed not to the instructor but to the students themselves. From the campus-speech week:
Across 42 discussions you pushed past easy slogans and did the harder work of naming who decides and where exactly the line falls. The strongest exchanges came on mascots, where many of you converged on deferring to the affected tribe while distinguishing slurs from tribe-specific names, and on hate speech, where the split between preventing harm and protecting against overbroad rules produced genuine, productive disagreement. Keep pressing on the authority question we kept circling: what should happen when the group a rule is meant to protect doesn't speak with one voice?
A joint interview with both instructors, on how the two courses came to be linked and what each would change next time, is planned and will be added to this showcase when it's complete.
Coda
Strangers by design, on purpose
Two instructors could have run their courses separately and lost nothing on paper — a syllabus doesn't require a partner school four states away. What they arranged instead gave eighteen students a semester in which arguing with a stranger from another campus was close to the default experience, and gave ninety-one students the same thing far less often. The two classes did not end up mirroring each other. The seminar's near-total completion rate and its instructor's account of a transformed semester sit at one end of this record; a sharper, more mixed set of reactions from parts of the larger class sit at the other, and this page has tried to show both.
What the fourteen weeks hold in common, across every student profiled here and the dozens more in the "more" drawers, is a record of people getting more exact — about what a statement actually claims, about what evidence would change their mind, about the precise word a disagreement turns on. Guide asked some version of "what do you actually mean by that" repeatedly across those fourteen weeks, on Gaza and gender and inheritance tax alike, and in everything we read, it never argued for a side. That is the distinction we mean by facilitative AI, as opposed to the generative kind: it did not produce the arguments, or the conclusions. It pushed on both sides evenly, and left the reasoning to the two people doing it.
Up next
The material behind this showcase, and how to run your own.
More showcases
Other courses followed through a full Sway implementation, from first assignment to last.
See all showcases →Instructor reports
Browse real (deidentified) reports — the format this showcase is built from.
Browse the reports →Run this in your course
Sway is free for instructors and students. One assignment, or a semester linked across two campuses — the design is yours.
Get started →