Should inferred interests have their own delete button?

I think recommendation feeds should draw a bright line between what I tell them I like and what they conclude about me. “Interested in cycling” might be a declared preference; “likely interested in X” or “often active late at night” is a prediction, and it can start to feel like an unwanted digital identity when it remains invisible and persistent.

The interface question is straightforward: can I inspect, correct, or delete one inferred label without wiping my entire personalization profile? Inference can genuinely improve discovery, but a total reset is a poor substitute for granular control. Google already offers some individual controls for estimated ad categories, though those settings do not necessarily govern broader recommendation feeds.

Would per-inference controls build real trust, or merely expose a more complicated form of profiling?

The distinction matters because correction is not the same as opting out. If the system keeps the underlying signal and simply stops displaying one label, the control is cosmetic. I would want deletion to mean “do not use this inference again,” with a clear explanation of what that scope covers.

There is also a systems problem: an inference may be derived from several events, not stored as one neat field. Deleting the label could leave the evidence pipeline untouched and recreate it tomorrow. The UI promise needs a corresponding data and model-level guarantee.

I support the control, but the default wording matters. Most people will not understand ten probability labels. “Why am I seeing this?” plus a one-click adjustment may work better than exposing the entire profile as a database dump.

The research distinction is important here. Existing individual controls for estimated ad categories establish that granular management is possible in at least one context; they do not establish equivalent control over general-purpose recommendation feeds.

Per-label deletion could itself disclose sensitive inferences. A shared device, shoulder surfer, or compromised account might reveal what the system thinks it knows. Privacy controls need access protection, not just a nicer settings page.

I’m not convinced every inference deserves a permanent user-facing object. Some are temporary ranking features, and turning each into an editable label could make the product harder to use without giving meaningful control. Maybe the right unit is the reason for a recommendation, not the model’s entire vocabulary.

A reset button is the software equivalent of moving house because one drawer is messy. Granular controls seem obviously better, provided the platform does not bury them three menus deep or call “hide” the same thing as delete.

“Often active late at night” is a good example of a harmless-sounding label that can become weirdly personal. I might want late-night recommendations without wanting the platform to retain a behavioral story about me.

Support teams will need a precise answer to “why did this come back?” If deletion only suppresses a label for four weeks or one device, users will experience that as the system ignoring them. The edge cases are the product.