Maya Chen: Nathan, hey — tell me you got some sleep this week, because I did not.
Dr. Nathan Hayes: Marginally. Conference prep. You know how it goes.
Maya Chen: Maya Chen: I was — wait, actually, I want to tell you what happened Tuesday morning, because it's sort of the whole episode in miniature. I'm on my phone at like 7:45, and I notice I'm angry. Not about anything that happened to me. Just... angry. And I'm scrolling and I'm engaging and at some point I put the phone down and I thought — who taught me to do that?
Dr. Nathan Hayes: That's — yes. That question is exactly where I want to start.
Maya Chen: Because today we're getting into algorithmic feeds — and specifically, mm, not the abstract 'the internet made us angry' version of that story. The version where we have the receipts. Where we know what platforms actually knew and when.
Dr. Nathan Hayes: Dr. Nathan Hayes: Right, and the receipt — Facebook's internal researchers, in 2019, built fake accounts. Controlled test accounts. One set up to look conservative, one liberal. And they watched the algorithm radicalize them in real time. The conservative account got QAnon content. The liberal account got pushed toward increasingly extreme left content. Now — they documented this. Internally.
Maya Chen: And then kept the system running.
Dr. Nathan Hayes: Correct.
Maya Chen: That gap — between 'we observed this' and 'we chose to continue' — I mean, that's not a math problem. That's a decision someone made in a room.
Dr. Nathan Hayes: And the decision had a name. Facebook pivoted to what they called Meaningful Social Interactions — MSI — as their core ranking signal. The explicit logic was: content that generates comments and shares gets amplified over content that people just scroll past. Adam Mosseri, who was VP of the News Feed at the time, put this rationale on record publicly. It wasn't hidden.
Maya Chen: Which is... almost harder to sit with, right? Because it means it wasn't a secret algorithm doing something nobody understood. It was a stated design philosophy.
Dr. Nathan Hayes: So the question I'd actually pose is — is this a math problem that produced bad outcomes, or is it a decision problem where the math was working exactly as intended?
Dr. Nathan Hayes: Both, actually. And that's the part I want to make concrete, because I think the plain-language version of this gets lost. The algorithm is not trying to radicalize you. It is not trying to make you angry. It's doing one thing: predicting what you will interact with next. That's it. That's the whole job.
Maya Chen: And 'interact with' means what, exactly? Like — what counts?
Dr. Nathan Hayes: So the signal types are — likes, shares, comments, watch time. Those are the inputs the model is trained on. And here's where the mechanism gets interesting: outrage is the most efficient of those signals. Emotionally intense content — moral condemnation, intergroup hostility, fear — it generates all four simultaneously, and it generates them fast.
Maya Chen: Hm. Faster than, say, something that's accurate but calm.
Dr. Nathan Hayes: Correct. Accuracy is not in the objective function. Wellbeing is not in the objective function. Chronological order — not in it either. The model doesn't know those things exist. It only knows you clicked, or you didn't.
Maya Chen: Maya Chen: The analogy I keep reaching for — and you tell me if this is too loose — is like a bartender who notices you drink faster when there's a fight in the room. No judgment. No agenda. They just start pointing you toward the fights.
Dr. Nathan Hayes: That's — yes. That's actually quite precise. No malice. Pure pattern-matching.
Maya Chen: Which is — I mean, that's sort of worse, right? Because there's no one to argue with.
Dr. Nathan Hayes: Now, here's where I want to push into something that I think most people stop just before. James Druckman's team ran a randomized controlled experiment on Bluesky during the 2024 U.S. presidential election. They switched some users to a reverse-chronological feed — no optimization — and compared them to users on the standard engagement-optimized feed. The engagement-optimized feed demonstrably amplified intergroup, moralized, emotionally extreme content. That's the causal evidence. But the thing that actually stopped me — most users in that experiment didn't notice.
Maya Chen: Wait — didn't notice that their feed was different?
Dr. Nathan Hayes: The mechanism is invisible to the person inside it. You don't feel the selection happening. You just feel — whatever you feel.
Maya Chen: Which, mm — that's the surface explanation, right? 'The algorithm optimizes for engagement, outrage wins.' Most people stop there. But that's actually not the whole story, because — wait, no, I want to say this carefully — the user isn't just receiving the content. Something is happening to the user in the process.
Dr. Nathan Hayes: Exactly. And that's the layer underneath. The algorithm trains on your behavior, yes. But you are also being trained. That's where we need to go next.
Maya Chen: Because the way I keep thinking about it — it's not passive. It's not like you're just watching TV and the TV is watching back. You're actually... participating in your own training. Like, every time you respond to the chest-tightening post, you're casting a vote.
Dr. Nathan Hayes: Right. And we have the data on exactly what that looks like at scale. There's observational research on Twitter — more than 7,000 users, 12.7 million tweets — and what it found is that when someone's outrage expression gets positive social feedback, likes, reshares, it predicts more outrage expression in subsequent posts. That's a reinforcement learning schedule. Not a metaphor. That's operant conditioning running on human behavior.
Maya Chen: Twelve point seven million tweets.
Dr. Nathan Hayes: Yes. So the effect size is — it's not small. And the mechanism is clean: positive reinforcement increases the behavior it follows. That's Skinner. It's just running on social media infrastructure now.
Maya Chen: And the likes come from other people, not from the algorithm directly. So there's a whole... social layer on top of the algorithmic one? Like the algorithm surfaces your post, other people respond, and their response is the actual — wait, that means you're being trained by your network, not just the feed.
Dr. Nathan Hayes: Exactly — and the same research picks that up. Users also conform their outrage expression to the norms of their ideologically extreme networks. So it's not just the algorithm conditioning the behavior. It's social learning. The people around you, algorithmically selected to be like-minded, are also modeling the register.
Maya Chen: That's — hm. That compounds in a way that feels almost inescapable.
Dr. Nathan Hayes: Now — and I want to be precise here because this matters — the Twitter study measured outrage expression. Behavioral output. It did not measure internal anger. It did not measure changed beliefs. That step is not supported by the data.
Maya Chen: So the algorithm might not be making the person angrier inside. It's... teaching them to perform outrage more.
Dr. Nathan Hayes: That's the more defensible claim, yes.
Maya Chen: I mean — I don't know if that's actually more comforting. Because if you spend enough time performing something, at some point you start to... sort of become it? Or at least, the people around you only ever see the performed version.
Dr. Nathan Hayes: And we have causal evidence on the directional push. Piccardi et al. ran a browser-extension experiment on X — they used an LLM to re-rank feeds in real time, up-ranking hostile political content. Affective polarization increased. So the loop runs both ways: your behavior trains the algorithm, and the algorithm's output shapes your subsequent behavior. That's the bidirectionality that the observational data couldn't establish on its own.
Maya Chen: And the person inside that loop on Tuesday morning at 7:47 — they don't feel a loop. They just feel like themselves.
Dr. Nathan Hayes: And that's — now, that's the part I want to push on, because 'they didn't feel it' isn't the whole indictment. The BBC ran a whistleblower investigation in March 2026. More than a dozen insiders at Meta and TikTok, on the record. And what they described was — internal research showing outrage drove engagement, and that research feeding directly into product decisions to allow more borderline harmful content into feeds.
Maya Chen: More than a dozen.
Dr. Nathan Hayes: On the record. That's not a leak. That's a pattern.
Maya Chen: And Jeff Horwitz — Broken Code — he got to the internal documents. Engineers weren't just observing what the system was doing. They were explicitly directed to let more borderline content compete. That's the word that gets me. Directed.
Dr. Nathan Hayes: Directed. Yes. And the mechanism behind that directive — actually, this is the part that doesn't get enough attention — internal documents also revealed that downstream MSI was weighted too. So it wasn't just that your comment generated a signal. Content that provoked someone who then provoked someone else — that whole cascade fed back.
Maya Chen: Wait — so the outrage compounds downstream? It's not one step, it's—
Dr. Nathan Hayes: The amplification was layered. Yes. And the TikTok pressure is what tightened that. Meta saw TikTok's engagement numbers and — look, this is documented — moved toward more borderline content specifically to compete. Not less. More.
Maya Chen: So the competitive pressure pushed the whole ecosystem in the wrong direction. Like, TikTok didn't cause Meta to clean up. It caused Meta to... accelerate.
Dr. Nathan Hayes: Correct. Now — and I want to complicate this, because the picture isn't as clean as the whistleblower narrative suggests — the Nature study. Large-scale Facebook study. Exposure to like-minded sources was widespread, yes. But it did not itself cause political polarization. That's a meaningful distinction. Seeing extreme content and having your beliefs changed by it are different empirical questions.
Maya Chen: Hm. But — mm — I wonder if we're measuring the wrong outcome. Because maybe the change isn't cognitive. Maybe it's who you're willing to sit next to at Thanksgiving.
Dr. Nathan Hayes: That's — yes, that's a real possibility, and I'd want a study design that captures relational withdrawal rather than belief change to test it. But I can't assert it from the Nature data.
Maya Chen: Okay but even setting that aside — YouTube. The 'Not interested' button. They've actually tested how well that works, right? And it's — it doesn't work. Not really.
Dr. Nathan Hayes: Empirically limited efficacy is the finding. The user-facing control exists. The objective function underneath it doesn't change. If the model is still optimizing for engagement, telling it once that you're not interested is — it's like whispering into a room full of loudspeakers.
Maya Chen: So the button is real, but the agency it implies... sort of isn't. Which means the fix can't live at the user layer. The architecture has to change — but nobody with the directive power to change it has moved yet.
Maya Chen: And that's — I keep turning that over. The architecture has to change, but the incentive structure that would have to change first is the one that's... making everybody a lot of money. So. I mean, that's where we've landed, right? Algorithmic inevitability versus engineered choice — and the distinction only matters if someone with the power to engineer differently decides to. Which they haven't.
Dr. Nathan Hayes: And the causal direction is still genuinely unresolved. That's not a hedge — it's a real open question. Are the algorithms producing extremity, or are they following engagement signals that already-extreme users are generating? Because those are different problems with different solutions.
Maya Chen: Yeah. And I don't think we can answer that today. Maybe nobody can yet.
Dr. Nathan Hayes: Not with the current evidence base. No.
Maya Chen: Mm. The thing that won't leave me — that person at 7:47 on a Tuesday. They're not going to feel the Druckman study. They're not going to feel the MSI weighting or the downstream cascade or any of it. They're just going to feel like themselves, opening their phone. And understanding the mechanism — I genuinely don't know if that changes anything for them. That's the part I can't resolve.
Dr. Nathan Hayes: Neither can I. And I think that's probably the honest place to stop.
