I expect that adding instructions that attempt to undo training produces worse results than not including the overbroad generalization in the training in the first place. I think the author isn’t making a complaint they’re documenting a tradeoff.