Would you put holdouts of good features in the same bucket as negative testing? Let’s say someone develops an ML model that cuts down spam by 90%. It’s launched to all but 1% of users, so that you can consistently measure that the spam reduction is 90%, that the users are benefiting from it, and that the false positive rate is low.

From a different POV you’re intentionally worsening the experience for 1% of users. Is this a negative test? If so, is it morally different?