Maya Chen: I read a number this morning that I genuinely cannot get comfortable with.
Dr. Nathan Hayes: Which one.
Maya Chen: Forty percent. Meta's internal research found that outrage posts spread 40% faster once they weighted the angry emoji at five times a like. And the thing is — they saw it. They ran the study. Frances Haugen eventually walked that research out the door. And the formula kept running.
Dr. Nathan Hayes: Right. So — the mechanism there is straightforward. Engagement-based ranking means content is surfaced by predicted engagement signals. If you code anger to count five times more than approval in that signal, the system learns, very efficiently, to serve more things that make people angry.
Maya Chen: It's not — yeah, it's not a mirror. It's more like a mirror that secretly added a filter and then said nothing.
Dr. Nathan Hayes: That's the distinction that matters. The claim 'algorithms just reflect what users want' collapses the moment you look at the parameter choices. A parameter is a decision. Five times is a decision someone made.
Dr. Nathan Hayes: But here's where I want to steelman the defense for a second — because there is a version of this that sounds reasonable. The version where you say: look, we're just surfacing what people actually choose. No one forces the click. And that's not wrong, exactly. The problem is it conflates two different things — what people say they want versus what their behavior rewards. Users consistently report wanting accurate, high-quality information. Surveys, self-report, all of it points that direction. And then their dwell time, their shares — those reward outrage. Every time.
Maya Chen: So the algorithm isn't learning the wrong thing. It's learning the right thing about the wrong version of us.
Dr. Nathan Hayes: Precisely. And a reinforcement learning system doesn't care which version is more authentic. It optimizes on the signal it can measure.
Maya Chen: Okay, so — imagine Tuesday morning, 7:15 AM. Someone brand new to the app, coffee going cold, no watch history, no stated preferences. What does the system even have to go on?
Dr. Nathan Hayes: Only behavioral signals. And within days — not weeks, days — it has sorted that user toward the highest-engagement content it can find. Facebook actually ran this internally. The Carol's Journey study. They created a test account, politically moderate conservative, named Carol Smith. Within two days of joining, the recommender system had surfaced QAnon content. Extremist groups. Not because Carol asked. Because the system kept serving content fractionally more extreme than her current position — each step maximizing engagement — and that pathway moves fast.
Maya Chen: Two days. I — wait, I keep trying to let that land and I don't think I'm doing it justice. That's not gradual drift. That's — the speed is what should stop us cold.
Dr. Nathan Hayes: And the thing that compounds it — Meta already had this finding. Internal research. They knew the radicalization pathway was that fast, and the question of what they did with that knowledge is exactly where the stated versus revealed preference gap stops being abstract and becomes, now, a question of institutional decision-making.
Dr. Nathan Hayes: Now, here's where I want to complicate what we just said — because there's a counterargument that's actually worth taking seriously. Some behavioral science researchers looking at the data make the case that exposure to inflammatory content, genuinely false inflammatory content, is — it's not distributed evenly. It's concentrated. A small, highly motivated fringe accounts for a disproportionate share of the consumption.
Maya Chen: So the fringe gets radicalized fast, and everyone else is... what, fine?
Dr. Nathan Hayes: Not fine — but differently affected. The fringe concentration hypothesis doesn't say there's no harm. It says the causal story, algorithms as the primary engine of broad societal polarization, that claim runs ahead of what the evidence can actually support.
Maya Chen: Hmm. Though — wait, actually — I keep thinking about the Van Bavel PNAS Nexus study. Because that found something that isn't about radicalization exactly. It's about satisfaction. People engaging with outrage content reporting lower satisfaction with it afterward. That's — I mean, that's not a fringe problem. That's a design problem hitting everyone.
Dr. Nathan Hayes: That's the engagement-satisfaction gap. And yes — that's the 2025 finding, and it matters precisely because it's separable from radicalization. You don't have to end up in a QAnon group for the system to be making you worse off.
Maya Chen: Right. So narrower and deeper for some, and subtly corrosive for... everyone else.
Dr. Nathan Hayes: That's the honest shape of it, I think. And what makes it actionable — the Northwestern and University of Chicago registered-report study in Nature Human Behaviour, they actually compared engagement-based ranking against a reverse-chronological feed as a control. Algorithmic ranking amplified moralized, emotionally charged, intergroup content. The lever exists. Platforms know it exists. The tradeoff is — reducing engagement-based ranking measurably cuts polarization metrics, and it also cuts engagement. Real engagement numbers. That's the cost Meta, TikTok, YouTube, X have all declined to absorb.
Maya Chen: So they've seen the exit and just... haven't walked through it.
Maya Chen: And that's the part I can't — mm — I keep returning to this. The 2018 News Feed change. The internal research flagged harm. Frances Haugen brought documentation. And the tradeoff was never made. Not because nobody knew what the lever did. Because pulling it costs engagement numbers, which costs ad revenue, which costs the free service model the whole thing runs on. That's not a mystery to be solved. That's a choice being made continuously, every quarter.
Dr. Nathan Hayes: Which means the question isn't technical. Whether you can reduce engagement-based ranking to cut polarization — that's answered. The Northwestern and University of Chicago study answered it. The question is what regulatory or market pressure would actually force the cost to be absorbed. And nobody has answered that.
Maya Chen: We know what fixing it looks like. We just haven't decided we want to pay for it.
