Maya Chen: So I've been sitting with this thing all week and I want to just — drop it on you. Ready?
Dr. Nathan Hayes: Go.
Maya Chen: YouTube tried to reduce amplification of angry political content. The goal was less polarization. Engagement fell so sharply the experiment got shelved. The algorithm isn't broken — it's doing exactly what it was built to do. That's today.
Dr. Nathan Hayes: That's the crux, yes. And — I mean, this applies across all of them. Facebook, Meta, Twitter, TikTok, YouTube. The underlying principle is the same: engineer the feed to maximize engagement metrics. Likes, comments, shares, watch time. The algorithmic feedback loop means content that earns high engagement gets surfaced more, which generates more signals, which amplifies the emotional intensity further.
Maya Chen: Mm. So it's not that someone chose outrage —
Dr. Nathan Hayes: No. Outrage is what floats to the top when your only measuring instrument is the click. Accuracy is expensive to verify. A click is free. That's measurement poverty masquerading as a design philosophy. And what I want to get to — what are these systems actually detecting in human behavior, and why does that keep resolving to outrage?
Dr. Nathan Hayes: Think of it like a vending machine that learns which snacks you grab fastest. It doesn't know if you love them or if they're just the brightest color in the row. It restocks whatever moves quickest. That's the whole system. The algorithm has no model of your wellbeing — it has a model of your grab speed.
Maya Chen: And outrage is... the brightest color.
Dr. Nathan Hayes: Consistently. Now, William Brady and Molly Crockett at Yale actually quantified this — on Twitter, outrage expressions earn disproportionately more likes and engagement. So the loop runs in both directions. The algorithm surfaces more outrage, creators learn outrage language is rewarded, they write more of it, that earns more signals — I mean, it's not that one person decided this. The reinforcement cycle trains both sides simultaneously.
Maya Chen: Wait — it's training the creators too?
Dr. Nathan Hayes: Yes. That's the part people miss. And the neurocognitive layer is what locks it in — novelty and threat activate dopaminergic reward pathways. Extreme content is literally stickier than neutral information. Your brain prioritizes it. So the algorithm isn't fighting human nature here, it's... actually, no — it's exploiting the exact attentional architecture we evolved for detecting predators.
Maya Chen: Hmm. So our threat-detection wiring just — gets hijacked by a post about a politician.
Dr. Nathan Hayes: That's the attention capture part. But here's where I want to flag something — does dopamine explain what you *prefer*, or just what holds your gaze? Because those might be completely different mechanisms. The Twitter randomized experiment, nearly two million accounts, showed algorithmic feeds amplified political content across seven countries. Right-leaning content got disproportionate amplification in six of seven. That's the output. But whether the users *wanted* that diet — that's the question we haven't answered yet.
Maya Chen: Maya Chen: Smitha Milli's PNAS Nexus study actually measured this. Users are reacting to outrage, generating all these engagement signals, and then reporting lower satisfaction. The algorithm reads the click as 'I want this.' But the person is saying 'that made me feel worse.' Those are two completely different things happening at the same time.
Dr. Nathan Hayes: And the algorithm has no instrument for the second one. It cannot — I mean, satisfaction isn't computable from a click. So it's genuinely optimizing for a proxy that diverges from the actual preference. That's not a rounding error. That's a structural mismatch.
Maya Chen: Which brings me to Carol Smith.
Dr. Nathan Hayes: Tell me.
Maya Chen: Facebook's internal study — circa 2019 — they built a test account, Carol Smith, politically conservative, brand new. No expressed interest in conspiracy content, none. Within two days, the algorithm had recommended QAnon-adjacent groups to her. Two days. And what gets me is — that's not the algorithm reading a preference she had. That's the algorithm... sort of, creating one? Before she even had the chance to form it herself?
Dr. Nathan Hayes: That's the precise problem. The feedback loop doesn't wait for a stated preference — it infers trajectory from the first few engagement signals and accelerates toward the content type with the highest predicted engagement. Which, given what we know about the radicalization pathway — Exposure, Reinforcement, Group Integration — QAnon content scores extremely high. And the BBC whistleblower disclosures from March 2026 make it worse: more than a dozen insiders from Meta and TikTok confirmed the companies had internal research showing exactly this dynamic, and structurally, the products didn't change.
Maya Chen: So they knew. They had the Carol's Journey data, they had the satisfaction gap — and the feed stayed the same. I keep thinking about what that feels like from inside the feed. You're clicking because something grabbed you, not because you wanted it. And the algorithm is filing that as a vote.
Dr. Nathan Hayes: And that's the part I can't — I keep returning to the YouTube experiment. Not as a failure. As evidence. They reduced the outrage amplification, engagement dropped, they shelved it. Which means the reduction was working. They measured exactly the wrong kind of engagement going away and decided that was a problem to avoid rather than a signal to follow. The business model literally cannot absorb what fixing it would cost.
Maya Chen: So the question isn't really whether they could change the algorithm. It's... what kind of business model would even let them want to.
Dr. Nathan Hayes: Right. And somewhere — right now — a team at Meta or TikTok is looking at exactly this tradeoff. They have the satisfaction data, they have something like Carol's Journey on file, and they know what the numbers say.
Maya Chen: We started with YouTube shelving the experiment because engagement fell. That still just — sits there, doesn't it. Yeah. Thanks for thinking through this with me.
