Agreed. Expressed vs revealed preference is the devil in the details of building useful recommendation models that seems to always have this catch. It’s worsened by a hugely unaccounted for spread of experience among users. Anyone training models daily takes for granted the often subtle but unusual ways they’re careful to avoid miscommunicating preferences. Slow-scrolling past content without pausing to lessen interest capture weighting, for example, or using statistically associated concepts for steering. Neither is average user behavior, yet developers often assume it.
In the end, did the user actually tell you their preference? or did you simply attempt to give them one [at any cost]