How Algorithms Decide What You See Online

Open any app and the content feels uncannily personal: videos you can't stop watching, products you were just thinking about, news that confirms what you already believe. That is not coincidence. It is the output of recommendation algorithms — systems that decide what billions of people see, shaping culture, commerce, and politics one feed at a time.

These systems are among the most consequential technologies ever built. They are also widely misunderstood.

Collaborative filtering and the prediction game

The classic technique is collaborative filtering: if people like you enjoyed something, you probably will too. Streaming services pioneered it — 'viewers who watched X also watched Y' — and the idea scales beautifully. Modern systems go far deeper, using neural networks that ingest thousands of signals: what you watched, how long, what you skipped, the time of day, your device.

The objective function matters enormously. Most commercial recommenders optimize for engagement — watch time, clicks, session length. The algorithm does not care whether content is true, healthy, or diverse; it cares what keeps you scrolling. This single design choice explains an enormous amount about the modern internet, from rabbit holes to outrage cycles.

Engagement optimization and its consequences

Optimizing for engagement produces predictable side effects. Content that provokes strong emotion — outrage, awe, anxiety — performs well, so the system amplifies it. Niche interests get served brilliantly; shared cultural ground erodes as everyone's feed diverges. Creators, in turn, optimize for the algorithm, shaping what gets made in the first place.

Reformers propose alternatives: optimize for user-stated satisfaction, diversify feeds deliberately, give users real control over the objective. Some platforms now offer chronological options or 'not interested' controls — small levers against a powerful machine.

• Recommenders predict what keeps you engaged, using thousands of behavioral signals.
• Engagement optimization amplifies emotional, divisive, addictive content.
• User controls exist but remain weak against the default settings.

There are glimpses of a different model. Subscription-funded platforms, less dependent on engagement, can afford calmer feeds. Researchers prototype recommenders optimized for learning or well-being rather than watch time. The algorithm is not destiny — it is a design choice, and design choices can be changed when the incentives behind them change.

Recommendation algorithms are not neutral mirrors; they are editors with incentives. Understanding their incentives — engagement above all — is the first step to taking back control of your own attention.

Until the defaults change, the most powerful tool remains the oldest one: deciding, deliberately, what deserves your attention — before the algorithm decides for you.