This one, I think, is them just not willing to admit their mistake. I wouldn't be surprised if the experiment went like this: "Here are the inputs with only underline. In one of them the label is inside the input". In this case sure, users will better understand it's an input if the label is inside.