The Algorithm You Didn’t Choose: How Recommendation Engines Became Moral Gatekeepers
I noticed it first on YouTube. Not in any dramatic way — no conspiracy theory rabbit hole, no radicalization pipeline. Just a quiet shift. I watched one video essay about urban planning in Tokyo, and for the next three weeks, every recommendation assumed I was a transit nerd who wanted to see every subway map on the internet. Which, fine — I kind of am. But it got me thinking: who decided that this is what I should see next? And what values are baked into that decision?
The answer, as it turns out, is not a person. It’s an algorithm — a mathematical model trained on millions of other people’s behavior, optimized for a single metric: engagement. And that seemingly technical choice — optimize for engagement — is one of the most consequential ethical decisions of the internet age.
The Invisible Architecture of Attention
Every major platform you use — YouTube, TikTok, Instagram, Spotify, Netflix — has a recommendation engine at its core. These systems don’t just serve you content; they build a model of who you are based on what you click, how long you watch, what you skip, what you share. Then they feed you the content most likely to keep you doing those things.
This sounds neutral. It’s not. Every recommendation algorithm encodes a set of priorities. The YouTube algorithm, for instance, was reportedly optimized for “watch time” rather than “number of views” after a 2012 redesign. That single change shifted the entire platform toward longer, more emotionally engaging content — because emotional content keeps people watching. It’s the same logic that pushes TikTok’s For You page toward increasingly extreme or sensational material. The algorithm doesn’t know what “extreme” means. It just knows what holds your thumb still.
These are moral choices disguised as engineering decisions. As researchers from the Oxford Internet Institute have documented, recommendation systems routinely make trade-offs between engagement and user welfare, between personalization and diversity of exposure, without any transparent framework for why one value wins over another.
The Hidden Curriculum of the Feed
Think about what a recommendation algorithm teaches you, implicitly, about the world. If the algorithm shows you outrage-inducing political content because it keeps you scrolling, it’s not just recommending videos — it’s training your attention. It’s shaping what you believe is important, what you think other people care about, what you consider normal.
This is what media scholars call “agenda-setting,” and it’s been studied since the 1970s. But algorithmic agenda-setting is different from what newspapers or TV networks used to do. Those had editors — flawed humans with biases, sure, but humans who could be held accountable, interviewed, fired. Algorithms have no editor. They have a loss function and a training set. And the values embedded in that loss function are rarely disclosed.
A Spotify playlist curated by an algorithm might never show you a genre you haven’t already demonstrated a taste for. That’s great for keeping you subscribed. It’s terrible for actually expanding your musical horizons. The same dynamic on YouTube means a teenager who watches one video about depression might get a cascade of increasingly dark recommendations — not because anyone at the company decided that’s a good idea, but because the data suggests that sad people watch sad things, and the algorithm optimizes for what the data suggests you’ll click.
This isn’t hypothetical. The Wall Street Journal has documented how YouTube’s algorithm can steer users toward increasingly extreme content with as little as a single “recommended” video. The algorithm didn’t intend harm — it just doesn’t have the architecture for ethical intent.
Opacity by Design
One of the deepest problems with algorithmic gatekeeping is that nobody — not users, not regulators, not even the engineers who build these systems — fully understands how they work. Modern recommendation engines are neural networks with millions of parameters. They are not written in code that a human can read and reason about. They are trained. And what they learn is often surprising even to their creators.
This creates an accountability vacuum. When a platform’s algorithm amplifies misinformation during an election or radicalizes a user toward extremist content, who is responsible? The engineer who wrote the training pipeline? The product manager who set “engagement” as the optimization target? The executive who decided not to invest in content moderation? The answer, in practice, is nobody — because the system operates as a black box, and each actor can plausibly claim they didn’t know what the algorithm would do.
As I’ve written before about how gaming communities develop their own moral codes, the rules that govern a space are rarely neutral. They reflect the values of the people — or the systems — that create them. An algorithm optimized for engagement is a rule system that values attention above all else. That’s a moral stance, even if nobody calls it one.
Who Gets to Be Seen?
Recommendation engines also function as gatekeepers for visibility. The TikTok algorithm, the Instagram Explore page, the YouTube suggested sidebar — these determine what gets seen and what disappears into the digital void. This has real consequences for creators, for small businesses, for activists trying to reach an audience.
The ethical question here is about fairness. The algorithm doesn’t consciously discriminate, but it can produce outcomes that look a lot like discrimination. Harvard’s Berkman Klein Center has documented how algorithmic systems can amplify racial and gender bias by learning from biased training data. A recommendation algorithm trained on user behavior inherits all of society’s existing prejudices — then amplifies them at scale.
If a platform’s algorithm systematically shows fewer recommendations for creators from marginalized backgrounds — not because of explicit prejudice, but because the training data reflects broader societal patterns — that’s still a form of gatekeeping. It’s just one without a gatekeeper. And that absence makes it harder to challenge, harder to fix, harder even to name.
This connects to a broader point about justice in digital spaces. In an earlier piece on how institutions fail to deliver equitable outcomes, I argued that fairness requires more than just the absence of explicit bias — it requires structures designed with equity in mind. The same principle applies to algorithms. A system that merely avoids conscious discrimination is not fair if its training data is marinated in historical injustice.
What Transparency Would Look Like
There’s growing momentum around algorithmic transparency. The European Union’s Digital Services Act requires large platforms to disclose information about their recommendation systems. Researchers are developing auditing frameworks for algorithmic systems. Some platforms now offer users some degree of control over their recommendations — you can tell YouTube you’re not interested in a topic, or reset your TikTok For You page.
These are steps in the right direction, but they’re not enough. Real transparency would mean platforms disclosing what their algorithms optimize for — not the source code, but the objective function. What metric drives the system? Engagement? Watch time? Time on site? Diversity of content? User satisfaction? Right now, we mostly have to guess.
It would also mean independent auditing by researchers who can study how these systems behave in the wild. The platform companies have resisted this, arguing that their algorithms are trade secrets. But when a trade secret causes demonstrable public harm — from election interference to youth mental health crises — the balance shifts. The public has a right to understand systems that shape public discourse at scale.
Biker subcultures developed elaborate rituals around who gets to wear certain patches and what they mean. That’s a form of gatekeeping — explicit, transparent, accountable. You know who decides, and you know the rules. Algorithmic gatekeeping is the opposite: invisible, unaccountable, and constantly shifting. Which system is more ethical? At least the bikers know who to argue with.
Toward Algorithmic Citizenship
I don’t think the answer is to abolish recommendation algorithms. They are useful. They help me find music I would never have discovered. They connect people with communities they didn’t know existed. They are not inherently bad — they are inherently consequential.
The ethical demand, then, is not to tear down the system but to demand that its builders take responsibility for what they’ve built. If your algorithm shapes what millions of people believe, what they buy, who they vote for, how they see the world — that is not a neutral technical achievement. It is an exercise of power. And power requires accountability.
We need a new kind of digital literacy — not just “don’t believe everything you see online,” but something deeper. Understanding that every recommendation is a value judgment. That the feed is not a mirror of your interests but a construction shaped by someone else’s priorities. That the algorithm is not a neutral guide but a gatekeeper with its own hidden ethics.
Until platforms are required to disclose what their algorithms optimize for — and until independent researchers can audit those systems — we are all navigating a city built by architects who refuse to show us the blueprints. We can see the streets. We can walk them. But we don’t know why they curve the way they do, or who decided that some destinations should be easier to reach than others.
The algorithm you didn’t choose is making choices about your attention every day. The question is whether you — and the rest of us — get any say in what those choices are.
What recommendation have you received lately that made you stop and think about what the algorithm assumed about you? I’d genuinely like to know.