Still inferring from the current context after personalization is off is undeclared personalization

What it is

The personalization switch is already off, and home still ranks by this moment’s location, hour, the device now playing, this minute’s query. The user thinks they left “according to me”; the system switched to “according to here and now” and kept estimating. Undeclared contextual personalization is using live context to rank differently by person and moment after the long-term profile is off, while still wearing an “off” badge.

Context is not a neutral default. A default is the speakable key. Context is another estimate.

Why it happens

Long-term profile and context features are often two input blocks on the same scoring pipe. The product turns off the user vector and leaves place, hour, device, current query, clicks in this session. For the user, “off” means no longer varying by person; for the model, the person still varies by time and place, only the vector became a session. Copy on the switch and system state misalign, and autonomy is discounted: people think they stand on a checkable default, and actually stand on an unnamed session profile.

Contextual inference can also touch sensitive attributes again: late night + place + what is now playing. Off should have cut that secondary use. If only long-term tags are cut and the session channel stays open, exit is half a cut.

Studying it

After long-term personalization is off, manipulate context: hour, network (home / work), device, tap a class in-session then refresh. Measure whether the list moves with it, and whether people still believe the switch took. Independent variables: whether context features still score, whether the UI declares “still ranked by time and place.” Dependent variables: list drift in the off state, belief that “not according to me” now holds.

A contrast should include a truly non-contextual default (pure time or a pure editorial slate). If long-term is off and the list still moves with place, while people think they are looking at site-wide popular, the declaration has failed.

Where it stops holding

The current query should of course affect search results; that is retrieval, and should be named retrieval, not “off personalization still working.” Safety filters (minors, geo-rights) are not taste estimates. Environmental signals such as weather or traffic, if used as content, should appear as a visible key (“by weather in your city”), not hide inside the off state. This entry is “still estimating from context after off.” It does not restate the default’s editoriality, and it does not restate inspect-as-prerequisite.

Applying it

• When long-term personalization goes off, turn off session-level taste estimation by default. To keep context, write a second sentence beside the switch — “will still rank by time/place” — and allow that sentence to be turned off too.
• In the off-state scoring pipe, drop place, hour, device, session clicks; leave only the declared key.
• Check: same account, personalization off, open in two places and two hours. If the list moves with them and the title still says “site-wide popular” or “latest,” contextual personalization is running undeclared. Drop context features from the pipe and compare again.

Related

• Same group: L6.10.1 Being able to see the profile is a prerequisite of being able to reset it · L6.10.2 Reset has to say what will be lost, or users will not touch it · L6.10.3 The default ranking after personalization is off is still an editorial choice and must be intelligible · L6.10.4 Deleting one record and resetting everything solve different problems and both must exist
• Nearby: L6.05 Turning Personalization Off · L6.06 Inferred Preferences and Their Limits · L6.12 Conflict Between Personalization and Predictability
• Search terms: undeclared contextual personalization · session inference after off · context features