A bit of punishment is a must. You learn from damage and hopefully as early as possible. Some never learn or only learn it the 1000th time. That's really bad but unfortunately can't be changed. Damage is rarely good for anything unless you learn something from it. Unfortunately, people who repeatedly do something stupid usually are. The damage they cause is then only harmful. It seems comforting to many that an AI does not yet manage to correctly weight analyzed training material on the first attempt. That's not bad at all, it's more expensive but fortunately it can still be changed.