Narrowing is produced by the objective itself; no one has to intend it

What it is

No editor ordered “lock this person in.” If the ranking objective is “predict whether this person will click,” range will shrink on its own. Objective-driven narrowing is a geometric property of a relevance objective — the next display is pushed toward the neighbourhood of observed preference — without hostility, and without clicks having to be written as a KPI on their own.

Click optimisation is one objective that is enough to narrow. The claim here is: swap in “predict rating” or “predict completion” and it still narrows, because they share the geometry of hugging the observed.

Why it happens

A relevance model places the user near past interactions in feature space and takes neighbours. Neighbours of neighbours are closer; the set’s diameter falls with iteration. That is the objective working, not operations working. Operations can talk diversity every day; if the loss that ships is still error on clicks or ratings, the gradient does not walk toward diversity.

Pariser’s filter bubble names the social consequence of that environment. Mechanically there is no need to assume someone wanted isolation: isolation is the fixed point of hugging the observed. Blaming a malicious editor looks in the wrong place — the objective should change, not another mail that says “please keep it diverse.”

Studying it

Freeze interface and operations; swap only the loss: predict click, predict rating, predict completion, relevance plus a diversity regulariser. After several rounds, measure set diameter, category entropy, author concentration. Independent variables: whether the loss hugs the observed, whether an orthogonal term exists. Dependent variables: slope of coverage across training rounds.

The contrast that matters is “pure objective, no one intervening” versus “the same objective plus an explicit diversity term.” If coverage falls under the pure objective, narrowing can be booked to the objective; no editorial intent need be assumed. The product story in which clicks are the only KPI is the previous group’s problem.

Where it stops holding

When the user’s query, filters, or subscriptions have already narrowed extremely, the model is executing a request, not secretly shrinking. A safety objective that deliberately compresses coverage is a different objective, not a side effect of relevance. When the item pool is already tiny, the diameter has nowhere to go. This entry argues that a relevance objective will shrink on its own. It does not treat diversity having to be its own objective, and it does not treat random injection.

Applying it

• Print the loss that actually ships. If it is relevance error only, admit coverage will fall by itself; “we meant no harm” is not a rebuttal.
• Before a relevance model ships, run multi-round simulations on frozen users and plot coverage slope. If the slope is negative, change the objective before talking operations.
• Check: same data, same interface, drop only the diversity term from the loss and retrain. Coverage down and relevance up means the objective is shrinking, not a person closing a gate.

Related

• Same group: L6.08.2 Diversity costs short-term clicks, so it has to be an independent objective · L6.08.3 Users cannot evaluate content they never saw; absence produces no visible feedback · L6.08.4 Random injection is not structured diversity; it only adds noise · L6.08.5 Diversity must sit on dimensions the user cares about; cross-category mix does not fix a single viewpoint
• Nearby: L6.02 Filter Bubbles · L6.04 Feedback Loops · L6.09 Feedback Loops and Preference Entrenchment
• Search terms: objective-driven narrowing · emergent filter bubble · relevance geometry