Why the feed shows what keeps you, not what you wanted

You open an application to check one item. Forty minutes later you close it having seen a hundred, none of which you asked for and most of which you enjoyed. The session is usually described as a failure of self-control. It is more accurately described as a solved optimisation problem.

What everyone sees

The everyday account is that recommendation systems learn your taste and return more of it. Feeds feel personal, so the objective is assumed to be preference matching, and a poor match is filed as a bug rather than as a different target being hit accurately.

What is actually happening

Production ranking systems are trained on signals that can be logged: watch time, dwell, click, completion, return within a day. Preference is not observable; behaviour is. Published descriptions of large recommenders state plainly that the model predicts expected watch time or interaction, a quantity that tracks satisfaction in the common case and separates from it in the tail.

Why it stays hidden

The divergence is invisible from inside the session, because the feed is never wrong in the moment: each item genuinely beats the alternative of closing the application. There is also no counterfactual on screen. You cannot see the session you would otherwise have had, so the only evidence of misalignment is a feeling afterwards — the one kind of evidence a metric cannot absorb.