can a feed make you dislike the other side a little less?

This is the main thing I work on. It started as a course project (CS692, fall 2025) in the HCI lab at UW-Madison, advised by Prof. Jacob Thebault-Spieker, and it has outgrown the course.

The question, in one breath: can a brief, personalized, narrative-based intervention inside a social-media-like environment produce a measurable drop in affective polarization and out-group animosity, even within a single session?

Affective polarization is the feeling part of politics. Not “I disagree with your tax plan” but “I don’t trust you, and I’d rather my kid not marry one of you.” People sort into an in-group and an out-group fast, on thin cues, and the sorting hardens with every dehumanizing thing the feed hands them. The brain doing the sorting is plastic. It is shaped by what it is fed, and that cuts both ways. I got that framing from popular science, David Eagleman mostly, not from doing neuroscience, and I hold it loosely.

I think of this as a computer science problem before a social science one. The recommender systems that surface the most tribal version of every argument are not laws of nature. They are code someone wrote to optimize a number. The same machinery could, in principle, be pointed at a different number. This project is a small, careful attempt to find out whether that is true, one feed and one story at a time.

The intervention itself is simple to describe. Show someone a first-person story from a member of their own side who changed their mind on an issue that matters to them. Not the other side arguing. Their side, reconsidering. Then measure whether anything moves.

the design

A lab-based, between-subjects experiment. Each participant lands in one of three arms:

1. Narrative intervention. Personalized first-person stories from in-group members who shifted their view on an issue the participant cares about.
2. Common-ground or information condition. Same issue, but the content points at shared facts or shared values instead of telling a story. This arm is optional. I would like to keep it, because it is what separates “a story moved them” from “new information moved them.”
3. Neutral control. Comparable content with no political charge.

Measures are taken before and after within each participant, so every person is their own baseline, and the arms are compared on the change.

The pitch of the project is that the content is not generic. Two things get matched.

The issue. From the pre-measures, each participant’s most polarized issue is the one where issue importance and perceived disagreement with the other side are both high. A story about immigration does nothing for someone who only gets heated about housing.

The narrator. The story comes from someone who shares a non-political identity with the participant: a major, a hometown, that kind of thing. The point is to make “one of us” land before the reconsideration does.

I use existing instruments rather than writing my own. They are cited here by name and the items belong to their authors, so none are reproduced.

• Affective Polarization Scale, after Iyengar, Sood and Lelkes (2012)
• feeling thermometers, in the ANES style
• Partisan Identity Strength Scale, after Huddy, Mason and Aaroe (2015)
• Actively Open-Minded Thinking scale, after Stanovich and West (1997)
• a behavioral partisan-cue task, so at least one measure is not self-report

Three hypotheses. H1: the narrative intervention reduces affective polarization relative to control. H2: it increases actively open-minded thinking and reduces partisan bias on the behavioral task. H3: the effects are larger when the story matches the participant’s salient issue and shared identity than when it does not.

H3 is the one I care about most. It is the one that says the personalization is doing work, rather than any nice story doing it.

the build

The part of this that is software, which is the part I can show.

The study runs inside a simulated social feed. It looks and scrolls like the real thing, but every post in it is chosen. Specific intervention posts go in at specific positions, and the short-term effect on the reader is measured right after. Nothing in it is live.

Underneath is a TypeScript browser extension that scores posts and injects content into a feed. It grew out of bridgefeed, the prototype that blurred bait on x and slipped in the calmer version of the other side’s argument. This one is the same idea with a study attached: the scoring decides what a post is doing, the injection decides what goes next to it.

The personalization logic takes the pre-measures, picks the participant’s most polarized issue, finds a shared non-political identity, and selects a narrator and story to fit. It is deliberately boring code. The interesting decisions are about what counts as a match, and those live in the design above, not in a function.

In the current arm the in-group narrators are AI-generated. Each one voices a position slightly shifted from where the participant’s side usually stands, a controlled scale shift rather than a conversion story. This is where the project touches my LLM work. It is also the part I am least comfortable with.

Two tradeoffs, named. A simulated feed buys experimental control and pays for it in ecological realism. Whatever moves in the lab moved in a feed that I built, not the one people live in. And AI-generated narratives raise two kinds of question. Validity: are the stories believable, and does the model’s own bias leak into what counts as a reasonable shift? Ethics: a tool that generates a persuasive in-group voice to move attitudes is, structurally, a manipulation tool, and calling it humanization does not make that go away. I do not have a tidy answer. The last section holds the untidy ones.

where it stands

Plainly: the design and the evaluation framework exist, the browser extension exists, and the formal lab study has not run yet. There are no results, no participant counts and no effect sizes. Anything on this site that reads like one is a mistake I want to hear about.

The original plan was in the wild. Deploy the extension to real people’s real feeds and watch for months. That is the version with ecological validity and the version I still think is the right long-term shape.

It is not the version I can run right now. A longitudinal deployment needs funding for participants and more time than a course and a bit can hold. So the current arm is a controlled lab study with recruited participants, a simulated feed, and AI-generated in-group narratives standing in for organic ones.

I want to be careful about how that reads. It is a control-versus-realism tradeoff, made on purpose. The lab arm can answer whether the mechanism moves anything at all in a single session, cleanly. The field arm would answer the question that matters more, but noisily and slowly. Doing the clean one first is a reasonable order, not a retreat, and the extension is shelved rather than dead.

What changed, in one place:

• from a longitudinal field deployment to a single-session lab study
• from organic in-group posts to AI-generated narrators with a controlled shift
• from watching real feeds to inserting posts into a simulated one
• unchanged: the question, the measures, the three hypotheses, and the extension underneath

what i don’t know

The list I keep next to the design. None of these are rhetorical.

Does a single session move anything durable? The lab arm can detect a shift right after exposure. It cannot tell me whether that shift survives the walk to the parking lot. The field study was supposed to answer this, and it is the reason I still want to run one.

Do AI-generated narrators stay believable without biasing the effect? If participants clock the stories as synthetic, the in-group cue collapses and I am measuring something else. If they do not, I still have to ask whether the model’s idea of a reasonable shift is quietly shaping what counts as one. I do not know how to fully separate the narrator’s effect from the narrative’s.

How would you personalize adaptively? Right now the matching is fixed rules. The obvious next step is something that learns, a contextual bandit or an RL policy that optimizes for lower polarization instead of engagement. Everything about that sentence is loaded. An optimizer aimed at attitudes is still an optimizer aimed at attitudes, whatever the sign, and I would want strict constraints on it before it touched a real person.

Who governs a tool that can shift attitudes? If this works even a little, it is a tool for moving people. That is already true of every feed, which is the argument for building it. It does not settle who gets to hold it, what they are allowed to optimize, or how anyone outside would know. I do not think the answer is “the person who built it.”

Is a simulated feed telling me about feeds, or about my simulation? Related to the first one, but distinct. The control that makes the study clean is the same thing that makes it artificial.

If you have thought about any of these harder than I have, I would like to hear it.