Maya Chen: Hey. You good?
Dr. Nathan Hayes: Yes, ready. You?
Maya Chen: Yeah, yeah — okay, I want to start with a number. One to five. That's it. That's the ratio that I — I keep turning over in my head and I can't let it go.
Dr. Nathan Hayes: The Facebook angry reaction weighting.
Maya Chen: Right. An angry reaction scores five times what a like scores. And — I mean, nobody announced that. It's just baked into the math. Which feels like... sort of a confession? Like the system is admitting what it actually values.
Dr. Nathan Hayes: Now, importantly — that ratio isn't arbitrary. The mechanism is that the algorithm measured what behavior those reactions actually predicted. Anger predicted clicks, shares, dwell time at five times the rate. So the weighting isn't a policy choice, it's an empirical output.
Maya Chen: Which is — wait, that's almost worse? Like it didn't decide to value anger. It discovered that we do.
Dr. Nathan Hayes: Correct. That is the core of what we call engagement-based ranking — a recommender system that orders content by predicting which items generate the most measurable interaction. Past behavior drives future predictions. The system learned from us.
Maya Chen: Mm. Imagine a friend who only ever texts you when something terrible has happened — a disaster, a scandal, a fight — because those are the only messages you open immediately. Every time. The algorithm ran that exact experiment, on billions of people, and it drew the same conclusion your friend would.
Dr. Nathan Hayes: That's a clean way to put it.
Maya Chen: And the thing that sort of cracked this open publicly — at least the most recent crack — was March 2026. The BBC, Marianna Spring, reported that more than a dozen whistleblowers from Meta and TikTok came forward saying that internal research had confirmed outrage fuelled engagement. And the companies knew. And the harmful content — violence, extremism, sexual blackmail — it kept increasing on the feeds anyway.
Maya Chen: Which brings me to the question I actually want to sit with. What exactly did the algorithm learn, and how did it learn it?
Dr. Nathan Hayes: What it learned is — okay, step one is simpler than people think. A user clicks on a story that triggers anger. That click is a training signal. Not a vote, not a preference statement — just a signal. The model updates: this type of content, this user, this context, higher predicted engagement.
Maya Chen: And that update happens immediately?
Dr. Nathan Hayes: Continuously. Every interaction is feeding back into the model in near real-time. So step two — the next content the system surfaces is slightly more provocative than the last. Not dramatically. Marginally. Because marginally more extreme content generated marginally more engagement. And step three is where it compounds — that marginal increase gets rewarded, model updates again, the floor shifts upward.
Maya Chen: That's — the floor shifts. That's the part that gets me. It's not a one-time choice, it's a ratchet.
Dr. Nathan Hayes: Right. That is the outrage amplification feedback loop, technically defined. And — now, importantly — the platform doesn't need to intend the escalation. The escalation is the mathematical output of optimizing for engagement. Intent is almost beside the point structurally.
Maya Chen: But wait — is that actually true? Because I keep wanting to push on the 'neutral math' framing. Like, the math is running inside a business model. And if the business model is selling attention minutes to advertisers — which it is — then maximizing engagement isn't a side effect, it's... the product?
Dr. Nathan Hayes: Yes. That is the mechanism. Engagement as proxy for ad revenue — the platform isn't selling content, it's selling the time you spend looking at the screen. Advertisers pay for that attention. So the financial incentive and the algorithmic incentive are — they're not parallel, they're the same incentive expressed in different terms.
Maya Chen: Which means it's not a bug they could patch. It's structural.
Dr. Nathan Hayes: Structurally, yes. And — this is where Jeff Horwitz's 'Broken Code' becomes important, the 2023 account excerpted in MIT Technology Review. Because what Horwitz documented wasn't that Facebook stumbled into this. The internal safety researchers at Meta flagged the loop. There was documented internal tension between the engagement optimization teams and the safety research teams. And the feed decisions — they kept going.
Maya Chen: So the awareness predates the harm. That's the part I — yeah. That's what I needed confirmed.
Dr. Nathan Hayes: And it's not just internal documentation. The Northwestern and University of Chicago researchers ran large-scale empirical experiments on algorithmic amplification of political content during actual elections. External confirmation. The low-credibility content amplification pattern they found — sensational or false claims generating stronger engagement than accurate reporting — that maps directly onto what the whistleblowers described from inside Meta.
Maya Chen: Okay, so the loop is real, it's documented from inside and outside — but now I'm wondering, what is the algorithm actually responding to in us? Like, is it anger specifically, or is it something underneath anger?
Dr. Nathan Hayes: Two things, actually. And this is where I want to separate them because they get collapsed constantly. Novelty and outrage. They feel like the same thing — both make you stop scrolling — but the underlying mechanism is distinct. Novelty is an information-scarcity signal. Your brain responds to the unexpected because unexpected things are, evolutionarily, the ones worth updating your model of the world for. That fires on anything surprising. Outrage is different — that's the threat and moral-violation system. Completely separate pathway.
Maya Chen: Wait — so when I open TikTok on a Tuesday morning and I'm four videos in and suddenly my pulse is up, those could be... two different things happening?
Dr. Nathan Hayes: Quite possibly, yes. The first video might grab you on novelty — something unexpected, a format you haven't seen. And then video three is actually triggering a moral-violation response. The algorithm doesn't distinguish. Both produced a watch-through, so both get rewarded.
Maya Chen: Hm. And it doesn't care which one got you there.
Dr. Nathan Hayes: Correct. The engagement signal is — it's agnostic to cause. Which matters for interventions, incidentally, because if you're trying to address the outrage pathway, a fix aimed at novelty bias won't touch it.
Maya Chen: Okay and — mm — this connects to something I want to name carefully because I think people use the terms interchangeably and they're not the same. Filter bubble versus echo chamber. Are those the same phenomenon or...?
Dr. Nathan Hayes: Different. A filter bubble is algorithmic — personalization narrows what content you even encounter, based on prior behavior. You don't see the contrary perspective because the system didn't show it to you. An echo chamber has a social layer on top — it's not just the algorithm, it's that your actual social network is also homogeneous. Homophily is the term. You choose to follow people who share your views, and the algorithm then amplifies within that already-narrow graph.
Maya Chen: So the echo chamber is — wait, actually — you're saying it's partly a choice? Like, the algorithm is working with material we handed it?
Dr. Nathan Hayes: That's the uncomfortable part. Some researchers — and this is where the radicalization narrative gets complicated — when they experimentally isolate the algorithm's effect specifically, the magnitude is smaller than the popular story suggests. User self-selection, existing social networks — those account for a substantial share. YouTube's rabbit hole, the algorithmic radicalization pathway, is real. But not everyone with identical algorithmic exposure ends up radicalized.
Maya Chen: That's — yeah, that destabilizes something. Because the story we've been telling is very clean. Algorithm bad, user victim.
Dr. Nathan Hayes: And it's not that the algorithm isn't doing something — it is. Affective content bias is documented, the amplification pattern on X, on YouTube, on Meta, it's consistent across platforms. But the individual outcome depends on what you brought to the feed before you opened it.
Maya Chen: Which means — oh, that reframes the whole question, doesn't it. It's not just what did the algorithm do to us, it's... what did we give it to work with. And then — who knew that? Who knew both things at once?
Dr. Nathan Hayes: The March 2026 BBC report. Marianna Spring. More than a dozen whistleblowers from Meta and TikTok. That's — now, that's the point where this stops being a mechanism question and becomes something else.
Maya Chen: Because they named the harm specifically.
Dr. Nathan Hayes: Violence, sexual blackmail, extremism — and the allegation isn't just that it happened. It's that internal research confirmed outrage fuelled engagement, and company decisions allowed that content to increase anyway. Those are two separate claims and I want to hold them apart for a second.
Maya Chen: Okay — what's the distinction you're drawing?
Dr. Nathan Hayes: Correlation versus deliberate intent. Showing that outrage drives engagement — that's a measurement. It doesn't by itself prove an executive looked at it and said, yes, let's cause harm. The counterfactual nobody examined is: what would a safe, engagement-optimized feed have even looked like? We don't have that comparison.
Maya Chen: I — mm. I hear that, though I wonder if the counterfactual is doing a lot of work for the platforms here. Because if you know the outcome and you keep optimizing harder — at some point, is that not functionally a choice? Like, the intent might be profit, but the downstream effect you've confirmed is harm.
Dr. Nathan Hayes: That is a reasonable position. I just want the evidence to carry the weight, not the narrative.
Maya Chen: Fair. But — wait — what about the user controls piece? Because platforms always say: users have tools. YouTube gives you 'Not interested,' 'Don't recommend this channel.' That's supposed to be the corrective.
Dr. Nathan Hayes: Right, and this is where the empirical studies are — actually, I'd say damning is not too strong. Those controls show limited and inconsistent practical effectiveness. You press 'Not interested' and the category deprioritizes marginally, sometimes not at all. The user experience is that the button works. The data says otherwise.
Maya Chen: That's — okay, that's enraging in a specific way.
Dr. Nathan Hayes: Because structurally — and this is the important part — those controls are a design feature sitting on top of an engagement-optimization engine. They cannot override the business model logic that funds the platform. The engine is optimizing for watch time sold to advertisers. A 'Not interested' tap is, mechanically, a much weaker signal than thirty seconds of dwell time on the video you claimed not to want.
Maya Chen: So the system is... literally outweighing your stated preference with your behavioral one.
Dr. Nathan Hayes: Every time. And that asymmetry is — I mean, it's not an accident of design. It's what the design is for.
Maya Chen: Which lands us exactly where I didn't want to land. Because if the controls don't work, and the internal research was there, and the whistleblowers are saying what they're saying — then what exactly is the user supposed to do with that?
Maya Chen: I think that's the thing I can't resolve. And I've been sitting with it the whole conversation. If this is structural — if the outrage amplification isn't a deliberate choice but a side effect of optimizing for engagement — then that should actually make fixing it simpler. You just change the metric. And yet Meta, TikTok, YouTube, X — none of them have done that. None.
Dr. Nathan Hayes: No. And that's — importantly, that gap is the thing. Because 'change the metric' sounds like a product decision. It isn't. The engagement-attention-advertising logic is the revenue foundation. You don't layer a fix on top of that. You'd have to abandon engagement-based ranking altogether.
Maya Chen: Which nobody's done.
Dr. Nathan Hayes: Which nobody has done. And the business model tension — user wellbeing on one side, advertiser-funded attention maximization on the other — it remains unresolved. Across all four platforms. That's not a characterization, that's just the current state.
Maya Chen: Mm. So the loop is documented. The harm is documented. The gap between knowing and changing — that's documented too, that's what the whistleblowers gave us in March 2026. And we're just... sitting inside that gap.
Dr. Nathan Hayes: That's where the evidence leaves us, yes.
Maya Chen: Whether 'structural side effect' is actually a morally different thing from 'deliberate choice' when you've seen the research, you've seen the numbers, and you've kept the engine running — that's what I can't settle. I don't know if that distinction holds.
Dr. Nathan Hayes: That is the question the evidence can document but not answer.
