What are you doing to evaluate models without such negligence?

When are we going to stop training the models to be so relentlessly persistent and start asking questions when there is ambiguity or it gets stuck?