Dr. Nathan Hayes: Maya, okay — before we even start, I want to flip the usual order. I'm handing you the contradiction first.
Maya Chen: Oh, this is new. Go.
Dr. Nathan Hayes: Twitter's own internal large-scale randomized experiment — approximately two million daily active accounts in a reverse-chronological control — found the ranking algorithm amplified right-leaning political content over left-leaning content in six of seven countries. Now, engagement optimization is the mechanism: platforms rank by predicting likes, shares, comments, watch time, not accuracy. So far the story holds. Then — 2020, Facebook and Instagram assign consenting users to chronological feeds. Algorithmic ranking is off. And polarization doesn't move.
Maya Chen: Hold on — substantial change in what people actually saw, and still nothing shifted?
Dr. Nathan Hayes: Substantial change in exposure. No significant effect on issue polarization. That's the finding.
Maya Chen: Mm, so we've been — I mean, the whole story we tell about algorithms is sort of built on the assumption that the mechanism and the harm are the same thing. And they might not be.
Dr. Nathan Hayes: Right. And here's the plain-language version of why. Think of it like a vending machine that learns what you keep pressing. It doesn't know what's good for you — it just knows what gets your hand moving. That's the whole engine. Platforms rank content by predicting engagement signals: likes, shares, comments, watch time. Not truth. Not balance. Just — what gets the hand moving.
Maya Chen: And the hand moves fastest when we're angry.
Dr. Nathan Hayes: Precisely. Researchers actually have a label for the content that performs best — IME content. Intergroup, Moralized, Emotional. Conflict between groups, moral framing, emotional charge. That combination empirically outperforms almost everything else for engagement. And it's not because engineers sat down and said 'let's amplify outrage.' The optimization function found it. Emergent property, not conspiracy.
Maya Chen: Oh — so the machine doesn't want you angry. It just... notices that angry gets the hand moving.
Dr. Nathan Hayes: And the psychological substrate underneath that is negativity bias — well-documented, evolutionary threat-detection. Negative information reliably triggers more reaction than neutral or positive. The algorithm didn't create that. It just, now — it exploits it at scale. Picture a Tuesday morning: you open your feed, an outrage post about an out-group has four hundred comments, a thoughtful explainer has twelve. The algorithm is watching that gap in real time.
Maya Chen: And then — wait, this is the part that actually unsettles me — it's not just what you see next. Your like on that outrage post is training the person who posted it. They get the hit, they post more. So the feedback loop runs both directions.
Dr. Nathan Hayes: That's the reinforcement mechanism — expressing moral outrage online gets behaviorally reinforced through likes, shares, replies, which makes users more likely to post outrage-laden content going forward. And the WeChat experiment with 2.1 million users adds another layer: algorithmically curated feeds drove significantly higher engagement with novel content than peer-sharing alone. So it's not just negativity. Algorithms are actively pushing novelty beyond what users would organically choose. The machine is pulling you somewhere you didn't quite decide to go.
Maya Chen: Which brings me to something I — okay, I've been sitting with the YouTube piece because it's the one that doesn't fit. YouTube is everywhere in this conversation. The radicalization pipeline. You watch one political video, the algorithm walks you toward something more extreme, and further, and further. That's the story. But then peer-reviewed audits come back and say... those fears are overstated?
Dr. Nathan Hayes: Overstated is the finding, yes. Multiple audits.
Maya Chen: So what — what's actually happening there? Because it's not that radicalization doesn't exist, it's just that YouTube isn't the mechanism we thought it was?
Dr. Nathan Hayes: Here's the distinction I want to draw. Amplification — measurable. Radicalization requires a full causal chain: content exposure, then belief change, then behavioral shift. The evidence for each step weakens as you move along the chain. And the audits themselves — this is the methodological problem — audit results are highly sensitive to configuration choices. Literally the same platform, different research design, you can get contradictory conclusions. So we're building a policy narrative on genuinely contested empirical ground.
Maya Chen: That's — yeah. That unsettles me more than the algorithm itself, sort of.
Dr. Nathan Hayes: And now here's what I didn't expect — users left entirely to their own preferences, no algorithm, choose less diverse content than the algorithm gives them. The echo chamber isn't just something platforms impose. The exposure is prevalent on Facebook, the mechanism is real, but the causal role in actually increasing polarization? Contested. People seem to find their way to the same narrowness with or without the recommendation engine.
Maya Chen: Wait — so we'd self-sort anyway. We're maybe narrating the algorithm as a villain because it's a legible villain, and the actual problem is... us, sort of choosing the same walls ourselves.
Dr. Nathan Hayes: That framing is — I mean, it's not wrong, but I want to be precise about what we're actually stuck on. Because the intentional-versus-emergent question isn't just philosophically interesting. It has real legal and moral weight. If a platform deliberately tuned for outrage, that's one kind of culpability. If the optimization function found outrage on its own — as an emergent property of rewarding engagement — that's genuinely harder to regulate. You can't subpoena an emergent property.
Maya Chen: So what are we even asking platforms to be responsible for? Effects we can measure — the amplification is measurable, the engagement–satisfaction gap is documented — but the downstream consequences we can't fully trace. That feels like... I don't know how you build accountability on that.
Dr. Nathan Hayes: I keep turning the engagement–satisfaction gap over. The algorithm gives you what makes you click — not what you'd report wanting. That's the documented divergence. And I can't tell you whether closing that gap requires intent on the platform's part, or just... someone deciding to measure for it. I genuinely don't know where that leaves us.
Maya Chen: Mm. Yeah. That's an honest place to stop, I think. I'm glad we didn't tidy it.
