Also, almost every A/B which results in a measurable impact in user response will have done so by positively vs. negatively impacting the users' emotions on the A/B legs of the study.

But more to the point, this is not even a fair characterization of what Facebook did. The made a tweak to their ranking algorithm which they deployed to 1/2500 of their users. Much later they discovered there was a extremely small but measurable impact to those users' engagement, and the types of words they included in their subsequent posts. They actually didn't know at all ahead of time what the effect would be, and in fact the effect they observed was the opposite of the widespread expectation.