That's just something they say to scare the children.
In any event, the point of a decent algorithm is that if the result isn't complying with the action, upcoming updates to the weights will fix it. Moreover, changes to the weight would be such that they optimize for maximum learning.
It is so weird seeing people preach for an obscure entity to do something so basic, and being shut down when asking for transparency. Today's AIs could write good model-development algorithms for tasks that are a hundred times more complicated.