The reason recommendations feel personal but arrive collectively

A recommendation appears that fits precisely, and the impression is of being understood. The computation underneath contains no model of you as an individual. It located accounts whose past behaviour overlaps with yours and offered what they consumed and you have not, which is a statement about a group rather than about a person.

What everyone sees

Personalisation is read as a profile: the system has assembled a description of your preferences and consults it. Some systems do hold attribute models. The dominant mechanism is comparative rather than descriptive, and it can produce accurate suggestions while holding no representation of why you might like anything.

What is actually happening

Resnick and colleagues introduced collaborative filtering by predicting a reader’s rating from the ratings of users who had agreed with them previously, requiring no information about the content itself. Sarwar and colleagues reformulated the same logic around items, computing similarity between items from the overlap in who consumed them, which scaled the approach without adding any model of the individual. In both designs the prediction is an average over a neighbourhood. Accuracy therefore rises with the number of people who resemble you, and personalisation is really the resolution at which the crowd has been sliced.

Why it stays hidden

The collective basis hides because the output is addressed to one person and arrives in a private context. There is no visible neighbourhood, no list of the accounts your suggestion was derived from, and no indication that the same item is being shown to thousands of others in the same slice. The experience of being known is produced by delivery, not by the computation.