Maya Chen: Sixty-two percent. That's the share of politically charged tweets that expressed anger when X's algorithm picked them. Versus 52% when you just — sort of — let the feed run in chronological order. Ten points. Which sounds like, okay, not huge. But that's applied across billions of impressions every single day.
Dr. Nathan Hayes: And the reverse-chronological feed is the control condition in that audit. That's the baseline — no ranking, just time order. So the gap is actually measurable. You can isolate what the algorithm is adding.
Maya Chen: Right. Which is — I mean, that's what makes this feel different from just, like, 'social media is bad.' It's a specific mechanism. Engagement optimization. Clicks, shares, watch time, comments — that's what Facebook, YouTube, TikTok, X are all actually ranking content by. Not accuracy. Not — whatever the opposite of outrage is.
Dr. Nathan Hayes: Not deliberative value, no.
Maya Chen: And high-arousal content — anger, fear, moral outrage — that reliably scores higher on those signals. So the algorithm isn't choosing hate. It's choosing what gets clicked. And those happen to be the same thing.
Dr. Nathan Hayes: That's the structural incentive. Advertising revenue is tied to time-on-platform. Nobody programmed animosity into the system. The system just rewards whatever keeps you there — and emotionally intense content does that more efficiently than anything else.
Dr. Nathan Hayes: Now, platforms will say — and this is the steelman — they're just serving revealed demand. You clicked it. You shared it. The algorithm learned from you.
Maya Chen: Right, 'we're just giving people what they want.'
Dr. Nathan Hayes: And that holds — until you look at what Kim, Buntain, and Ciampaglia actually found in 2026. Mixed-methods study, young adults. Participants were engaging with low-quality content they — they explicitly did not endorse. Like, when asked, they said they wanted accurate, high-quality information in their feeds. But the behavioral data? Implied something completely different.
Maya Chen: Wait — so they're clicking things they would actively tell you they don't want to be clicking?
Dr. Nathan Hayes: That's the stated versus revealed preference gap. And — importantly — this isn't unique to social media. There's TV news research from the 1980s showing viewers claimed to want substantive reporting but viewership data showed they watched sensationalism. We've known this for decades. What's different now is the industrialization of it.
Maya Chen: Hm. That's — yeah. Because in the 80s nobody had an algorithm learning your exact reflex patterns at scale and then — I mean, feeding them back to you a hundred times a day.
Dr. Nathan Hayes: Right. And the 67% figure matters here — language targeting out-groups increased sharing likelihood by 67%, was a strong predictor of angry reactions. That's from June 2021. So the algorithm isn't reflecting a considered preference. It's exploiting an attentional reflex. Which means — wait, does that mean users aren't actually choosing this?
Maya Chen: That's — yeah, that's the thing that keeps catching me. Because the feedback loop doesn't stay static, right? The algorithm learns that emotionally intense content gets engagement, so it surfaces more of it, and then — I mean, what was high-arousal last year is just... baseline now? The threshold keeps shifting upward.
Dr. Nathan Hayes: That's the amplification spiral. And it's compounded by collaborative filtering — the system isn't just learning from you. It's inferring your preferences from users who behave like you. So if the broader population skews toward outrage-coded content, that bias gets folded into your feed even before you've clicked anything.
Maya Chen: Wait — so it's not even purely reactive to what I did. It's also reactive to what people like me did.
Dr. Nathan Hayes: Correct. That's algorithmic bounded rationality — personalized recommendation systems actively narrow your exposure by reinforcing confirmation bias, negativity bias. It's not that you chose a filter bubble. The architecture builds one around your behavioral profile. Now — and this is where I want to be careful — that's the robust finding. What's not settled is the magnitude.
Maya Chen: Meaning — we know amplification happens, but we don't actually know how much of real-world polarization it's driving versus, I don't know, geographic sorting or partisan media or economic inequality.
Dr. Nathan Hayes: Exactly. 'Algorithms amplify outrage' — that's established. 'Algorithms are the primary driver of societal polarization' — the causal evidence for that specific claim is much weaker. The audit studies catch the directional signal. They can't close the gap between a 10-point anger amplification and what we actually see fracturing in communities.
Dr. Nathan Hayes: And that's where I keep getting stuck. Because forced diversity mechanisms — technically those exist. You can build a recommendation system that's required to surface a broader range of content types, not just whatever maximizes engagement. That's deployable. It's not a theoretical future thing. But the distance between 'technically deployable' and 'actually deployed' — I mean, that distance isn't an engineering problem. The advertising revenue model is tied to time-on-platform. If you sand down the outrage, you sand down the retention. So the fix, structurally, doesn't live inside the algorithm. It lives in whether the economic incentive changes.
Maya Chen: Which means it doesn't live with the user at all.
Dr. Nathan Hayes: No. If the incentive is structural — advertising revenue, time-on-platform, the whole architecture — then individual behavior change, media literacy campaigns... the effect size there is negligible. It's not that those things are worthless. It's that they're operating at the wrong level. The lever is policy or platform architecture. And neither of those is — I mean, neither is anywhere the person listening to this can actually reach.
