It's still algorithm-first. It's not "learning data" in the sense that human ratings are used to train the algorithm. It's more of a check step. So, for exampl

It's still algorithm-first. It's not "learning data" in the sense that human ratings are used to train the algorithm. It's more of a check step. So, for example, if humans think one website is vastly more (or less) authoritative than human raters, the engineers might dig in to see which aspects are causing the algorithm to evaluate it differently and, potentially, test tweaking the algorithm accordingly.