The Algorithm Knows You Better Than You Think

Here is a small, clarifying fact about how you consume information online: the platform you are using right now knows far more about what you actually want than you do.

Not what you say you want. Not what you click thumbs-up on or consciously seek out. What you linger on. What you return to. What you hate-watch at 11pm and feel vaguely ashamed of afterward. That behavioral residue—implicit, unreflective, deeply human—is the raw material that content recommendation algorithms are optimized against. And understanding that distinction is the starting point for a recent video from Crash Course that is ostensibly a fundraising pitch but lands, unexpectedly, as one of the more lucid explainers on algorithmic dynamics I've encountered in a while.

The video features Hank Green, one of Crash Course's co-founders, eating peanut M&Ms and raw garlic cloves while making a genuinely interesting argument. The garlic is a coin-flip gimmick to hold your attention—more on the irony of that in a moment.

The architecture of distraction

Green frames social media platforms, including YouTube, as complex systems: entities where behavior emerges from the interaction of components rather than from any single component's properties. This is a meaningful framing, not just jargon. It's why the oft-repeated claim that "the algorithm is designed to make you outraged" is simultaneously sort of right and subtly wrong.

Platform engineers don't sit in rooms designing outrage machines. What they design is optimization machinery—systems that maximize engagement signals, which turn out to correlate heavily with negative emotion because negative emotion is remarkably effective at holding human attention. The outrage isn't designed. It's selected for. That distinction matters, because it means the problem isn't one bad actor with bad intentions; it's an emergent property of architectures that are working exactly as specified.

Green puts it plainly: algorithms are "plugged into our most under-baked lizard brain reactions, giving us more and more of that video you hate-watched in the moment and then felt gross about later, and less and less of what the best versions of ourselves actually want to see."

The phrase "best versions of ourselves" is doing a lot of work there, and it's worth sitting with. It implies a gap between what we would choose under conditions of reflection and what we actually choose under conditions of infinite scroll. That gap—between considered preference and revealed preference—is where algorithmic optimization lives. Behavioral economists have studied this gap extensively. The algorithm has simply found a way to monetize it at scale.

The tuberculosis problem

Green offers an anecdote that illustrates the pressure this creates for creators more vividly than any abstract argument could. His brother John made a Vlogbrothers video titled "We Know How to End Tuberculosis." Tuberculosis, for reference, kills approximately 1.25 million people per year—more than any other single infectious agent. The video was, by the channel's standards, underperforming badly.

So Hank retitled it: "Elon Cancelled This, Let's Do It Anyway."

"It went from being our worst performing video of the month to the best performing video of the month," he says, "which was good for educating the world about tuberculosis but also I still felt kind of bad about it."

That parenthetical discomfort is the heart of the problem. The information in the video didn't change. The argument didn't change. What changed was the emotional valence of the packaging—from neutral informational framing to conflict-and-resentment framing. And the algorithm, faithfully doing its job, rewarded the latter enormously.

This is not a hypothetical pressure that creators talk about in theory. It is a documented, measurable phenomenon that shapes editorial decisions every day, across every platform, at every scale. And it's not limited to YouTube. Decades of research in science communication have shown that health information framed around threat and loss consistently outperforms information framed around opportunity and benefit—a pattern that long predates algorithmic amplification. The algorithm didn't create this vulnerability in human cognition; it found it, mapped it, and built an expressway through it.

Green also gestures at the misinformation dimension: "Some studies have even shown that lies spread more quickly than true information on social media precisely because they seem novel or surprising." This is a reference to well-established research—most notably a 2018 MIT study in Science that found false news stories on Twitter were 70% more likely to be retweeted than true stories, propagating faster and reaching more people. Novelty captures attention. Lies are, almost by definition, more novel than accurate descriptions of a complicated reality.

The cost of not playing

What makes Crash Course's position interesting, and worth examining rather than simply accepting, is the specificity of their claim: that they have largely refused to optimize for outrage, and that 143 million views in the past year suggests this strategy is sustainable.

Green is careful to acknowledge the partial contradiction embedded in the video itself—a coin-flip gimmick is, at its core, a variable-reward mechanism designed to maintain viewer engagement, which is a technique the algorithm would approve of heartily. He admits as much: "I am kind of doing that right now." The self-awareness is genuine, but it also points to something real: you cannot fully exit the attention economy while still operating within it. The question is always one of degree.

The more substantive point is about content selection rather than packaging. "We're not choosing what stories to tell based on what's going to make the algorithm happy," Green says. "We're choosing to teach what needs to be taught." Last year that meant series on Native American history, Latin American literature, sex education, and scientific thinking—subjects that are, as Green notes, "under-taught and under-funded" precisely because they don't generate outsized engagement on their own merits.

The economics of this are real and worth understanding. Each Crash Course video involves more than 20 people—subject matter experts, writers, editors, animators, producers. That is a substantial production cost for content that deliberately avoids the virality strategies that might otherwise offset it. Crash Course's solution is direct community funding: a coin program (the 2026 version is currently available) through which viewers can fund learner outreach at various scales, supplemented by Patreon. The organization's parent company Complexly recently converted to nonprofit status, which changes both its structural incentives and its funding possibilities.

Whether community-funded educational media is a scalable alternative to algorithmic incentives, or a niche model that works for established properties with loyal audiences but offers little guidance for new entrants, is a genuine open question. Crash Course built its audience largely before the current recommendation landscape solidified. A new educational channel launching today would face a different environment.

What the algorithm can't tell you

There is a version of the algorithmic critique that tips into fatalism—the idea that human attention is so thoroughly captured by these systems that the project of deliberate, reflective information consumption is essentially lost. Green doesn't go there, and I think he's right not to.

The 143 million view figure is relevant here, not as a boast but as evidence. It doesn't prove that most viewers are making fully deliberate, reflective choices when they watch Crash Course—people find content through recommendations all the time without any particular intentionality. But it does suggest that demand for substantive educational content is not negligible, even in an attention economy optimized against it.

What the algorithm optimizes for is revealed preference—what you actually do. What it cannot capture is considered preference—what you would choose if you stopped and thought about it. That gap exists. It is, arguably, what educational institutions have always tried to address: the space between what is easy to attend to and what is worth attending to.

The question Crash Course is implicitly posing, underneath the coin-flip gimmick and the fundraising ask, is whether that gap can be maintained and widened, or whether algorithmic systems gradually collapse it—training us, over time, to not even notice it's there.

---

By Priya Sharma, Science & Health Correspondent