The feed is a prediction machine

What is actually happening when you scroll

Every time a feed loads, three things happen in under a second.

First, candidate generation: from millions of possible items, a fast and rough process pulls a few hundred that might suit you. Second, scoring: a much more expensive model predicts, for each candidate, how likely you are to do each of several things: watch to the end, rewatch, like, comment, share, or close the app. Third, ranking: those predicted probabilities are combined by a weighted formula, and the order you see is the result.

The weights in that final formula are a business decision, not a learned one. Somebody chose how much a predicted share is worth relative to a predicted long watch. When a platform announces it is “showing more content from friends”, that is usually those weights being changed, not a new model.

This structure is published. YouTube described the two-stage design in a 2016 paper; Instagram has published lists of its ranking signals; Twitter released a version of its ranking code in 2023. The details differ, the shape does not.

What it is optimising for

This is the sentence to keep. A feed does not predict what is good for you, or true, or important. It predicts what you will do next.

Everything else follows. Content that reliably provokes a response is promoted because provocation predicts response. Content that is quietly useful and produces no reaction is demoted, not out of malice, but because the model has no column for “was quietly useful”.

The strongest signal on most video platforms is not the like button. It is watch time, and specifically whether you finished and whether you watched again. That is why a short video you rewatch by accident moves your feed more than a video you deliberately liked. If you want to steer a feed, the effective lever is what you actually watch to the end, not what you tap.

Why short video learns you faster

A fifteen-second clip gives the system a complete signal every fifteen seconds. A forty-minute video gives one signal in forty minutes. Feed the same learning machinery a hundred times more feedback per hour and it converges on your preferences far quicker.

That is most of why short-form feeds feel uncannily accurate within a day, while a film recommendation service still suggests things you would never watch after a year. It is not a smarter model. It is a faster loop.

Where the honest uncertainty is

The filter-bubble story (that these systems trap people in a narrowing tunnel of agreement) is far less settled than headlines suggest.

Several large studies, including a set of 2023 experiments run with Meta’s cooperation during a US election, found that switching people to a chronological feed changed what they saw a great deal and changed their political attitudes very little over the period measured. Other work finds real amplification effects. Researchers disagree, the experiments are hard, and the platforms control the data.

So the fair statement is: these systems demonstrably shape attention, and their effect on belief is contested. Anyone who states either side as established fact is going beyond the evidence.

What you can actually do

Unlike most systems in this block, this one has real controls, and they are free.

• Watch and search history is what the model uses. On Google and YouTube it lives at myactivity.google.com, where you can view it, delete items, pause collection, or set automatic deletion after three months.
• ”Not interested” and “Don’t recommend this channel” feed the model a negative signal directly. They work, though slowly, because one signal competes with thousands.
• Logged-out or incognito viewing gives the system much less to work with, at the cost of worse recommendations, which is exactly the trade being made.
• On most platforms there is a following-only or chronological view buried in the settings. It is usually not the default, and usually reverts.

None of this is about resisting a manipulation. It is about knowing that a system is predicting your behaviour from your behaviour, and that the only input you control is the behaviour.