Worse in a really insidious way too. Optimizing based on engagement for instance could be maxxing the amount of time people spend using a service, while minning unseen variables people care about like their emotional state while engaged. It’s sort of an inevitable thing that optimizers do to things they don’t measure, and it’s an extremely difficult problem to put everything people care about into the equation. Like for instance, think of how horrible a polynomial fit gets for a function just outside the window you are fitting as you add more terms.