I think the main issue with these metrics, which you implicitly highlight, Is that they are not a one size fits all approach. In fact, they are often treated, at least casually, like they are some kind of model fit like an r squared value. Which is maybe a good description narrowly constrained to the task or set of tasks they are being evaluated on for the metric. But the complexity of the user experience combined with the poor sample rate that a person can individually experience leads to conclusions like these. And they are perfectly valid conclusions. If the model doesn’t work for you, why use it? But it also suggests that personal experience cannot be used to decide if the model performs in aggregate well or not. But this doesn’t matter to the individual user or problem space. Because they should of course use whatever works best for them.